2d:Td25,# One campaign, and the rule that runs through it You're in the campaign studio. **What you're doing this lesson:** run one fall relaunch through a check before anyone approves it. You do not need to write a script today. Demo code, if it appears later, is optional. You're the marketer at Fieldnote, a coffee-subscription brand. Fall relaunch in three weeks: a new subscription tier, then a two-week push across email, social, and paid. AI will draft every piece of copy. That part is settled. Nobody is going to pay you to type first drafts by hand. What they will pay you for is this: **nothing publishes without a named approver, and nothing gets in front of an approver that hasn't passed its checks.** Copy has to sound like Fieldnote before Dana sees it. Calendar rows need a yes before they hit the scheduler. Report numbers have to come from the export before anyone upstairs quotes them. The AI drafts; you check; a human named in the file approves. Three distinct jobs, and the middle one is yours. The CMO Survey's Spring 2026 wave puts AI in 24.2% of marketing activities and climbing. The drafting already moved. The checking mostly didn't. ## The brand book problem Every brand has voice guidelines. Almost none of them are checkable. "Warm but confident." "Punchy." "Sounds like a knowledgeable friend." You cannot test whether a sentence is warm — and neither can the model. That's why pasting the brand PDF into a prompt produces copy that gestures at your voice while shipping "game-changing" in line one. The fix is to split the guidelines into two piles: - **Testable rules.** "Sentences under 20 words." "No exclamation points." "One idea per line." You can check every one of these. - **Banned phrases.** The cliches, the superlatives legal can't back, the competitor names. A list. If the phrase is in the draft, it fails. A real **voice pack** (the short list of measurable rules a script can check) looks like this: three testable rules, four banned phrases, a named owner (Dana), a version, a date. You'll see one in the next steps — skip the demo if you want. When Dana tightens a rule in August, the version bumps to v4 and every check downstream picks it up. That's the difference between a voice pack and a PDF nobody opens: the pack is *attached to the campaign*, not filed next to it. ## Why write the pack down, not just paste it in the prompt You'll still put the pack in the prompt — the model drafts better with it. But the prompt is a request, not a check. The model complies most of the time, and "most of the time" across 40 variants means three or four violations sail through per batch. After the model drafts, the same pack flags every violation mechanically. Same source of truth, two jobs — brief the model going in, check the model coming out. ## Where this lesson goes Next you'll watch a linter catch a bad draft, then tell banned phrases from mood words, then mark PASS/FAIL on a batch and build Dana's on-voice queue — the exact pass that runs before anything reaches approval. By the end, "on-voice" stops being a vibe and becomes a list of clear ship/hold calls. One placement note now, paid off in lesson 3: in a live stack this pass isn't a file you remember to run. It's a step between the model and the approval queue, so every batch walks through it whether anyone is watching or not.2e:Tac5,# The linter: fifteen lines that replace an argument You know the review meeting where someone says "this doesn't sound like us" and someone else says "it sounds fine to me" and twenty minutes evaporate? A linter ends that meeting. Either the draft contains "best-in-class" or it doesn't. Either the sentence runs 27 words or it doesn't. The conversation moves from taste to fact, and fact takes four seconds. **Code role:** optional runnable example — purpose is to show how a pack becomes a mechanical check in the AI draft workflow. You do not need to write this from scratch. Run the editor. The whole check is: 1. **Lowercase the draft once.** Banned phrases are stored lowercase; `"Revolutionary"` and `"revolutionary"` are the same sin. Case-insensitive matching costs one `.lower()` call. 2. **Loop the banned list, substring match.** `phrase in text` — Python's `in` on strings is a substring check. Each hit appends a labeled violation, `banned: revolutionary`, so the report says *what* broke, not just that something did. 3. **Word-count check.** `len(draft.split()) > PACK["max_words"]`. Crude twice over — `split()` counts tokens, not typographer's words, and it measures the whole draft against what the pack states as a per-sentence rule. That's fine here because this draft *is* one sentence; body copy would need a split into sentences first. But crude and consistent beats precise and manual. Same yardstick every draft, every time. Note what the violations *are*: strings, in a list. Not print statements buried in the loop, not a boolean `looks_ok`. A list of labeled problems is something the pipeline can count (`len(problems) == 0` means pass), forward (attach them to the variant that failed, send it back for a redraft), and log (which rules fail most tells Dana which rules the prompt needs to state harder). ## What the linter doesn't catch — on purpose "One idea per line" is in the pack but not in this script. Some rules stay human: the linter can't tell whether a sentence carries one idea or three. That's fine. The point was never full automation — it's that every *mechanically checkable* rule gets checked mechanically, so the human reviewer spends her attention on the rules that actually need a human. Dana shouldn't burn her review pass counting words. She should burn it on "is this true, and is it us?" The division of labor, explicitly: the script catches banned phrases and length. Dana catches meaning. The model catches nothing — it's the thing being checked. Next step: a draft goes through this exact check, and you call the violations before the tool does. The code above is a read-only example of what a teammate might automate — you pass by naming the hits, not by authoring the loop.2f:Taa6,# Variants are records, not paragraphs The old way to review AI copy: the model dumps twelve options into a doc, you scroll, you copy-paste the least bad one, and three weeks later nobody can say which subject line ran on which send or where it came from. The doc was write-only memory. The studio way: every variant is a record with a schema. **Code role:** read-only example of a variant record your AI workflow can lint and approve. Run the editor — that's one variant. The fields do real jobs: - **`id` and `channel` and `slot`** — so "E3 won the subject-line test" is a sentence with a referent. When performance data comes back in lesson 3, it joins on these fields. - **`text` and `cta`** — the actual copy, and the ask. `"cta": "none"` is a deliberate value, not a missing one: Fieldnote's pack says a CTA appears only when it earns its place, and the schema makes "no CTA" an explicit choice someone made. - **`source`** — `ai-draft`, `ai-edited`, or `human`. This field looks bureaucratic until the day someone asks what was AI-written in last quarter's campaign — for a rights question, a client dispute, a postmortem — and you can answer with a filter instead of an archaeology dig. - **`lint`** — the linter writes its violations here. Empty list means the variant may proceed to review. Non-empty means it goes back for a redraft, *with the reasons attached*. - **`approved_by`** — empty until a named human fills it. No script in this chapter ever writes to this field. That's not a technical limitation; it's the point. ## Why schema beats prose at batch volume Ask the model for "some subject lines" and you get prose to scroll. Ask it to return variants in this schema — and models are reliable at filling a stated JSON shape — and you get data to *process*: lint all twelve in a loop, sort the survivors by slot, count how many the pack rejected, and hand Dana a clean queue instead of a doc. The schema is what turns "review the copy" from an afternoon into a pass. Batch volume is where this pays. Writer's own customer story about the retailer Adore Me claims a 20-hour monthly product-description batch fell to 20 minutes with drafting agents — a vendor's number about a vendor's client, so treat the figure as marketing. But the shape of the claim is the believable part: batch drafting is where marketing time goes, and batch drafting without a batch *checking* step is how twelve variants of "game-changing" reach your send tool. You've got the pack rules from earlier. Next you'll mark four drafts PASS/FAIL against the pack, then — checkpoint — triage a five-variant batch into Dana's on-voice queue. Optional helpers can compare your marks; judgment is the unlock.30:T62f,PACK = { "banned": ["game-changing", "best-in-class", "revolutionary"], "max_words": 12, # subject-line cap — tighter than the 20-word body rule } drafts = [ ("V1", "Fresh beans at your door every Tuesday"), ("V2", "The game-changing subscription your mornings deserve"), ("V3", "We roast small batches weekly and ship them straight to your kitchen door for you"), ("V4", "Skip a week anytime, no questions"), ] # --- complete helper (optional tool — do not rewrite) --- # Purpose: same voice check a teammate might automate later. # Walkthrough: lowercase once → scan banned list → word-count check → list of labels. # Gloss: `phrase in lowered` means "this banned phrase appears somewhere in the draft." def lint_copy(text, pack): problems = [] lowered = text.lower() for phrase in pack["banned"]: if phrase in lowered: problems.append(f"banned: {phrase}") if len(text.split()) > pack["max_words"]: problems.append("too long") return problems # === YOUR JUDGMENT — edit only this block === # result: "PASS" or "FAIL" # violations: list of labels (empty when PASS) verdicts = { "V1": {"result": "PASS", "violations": []}, "V2": {"result": "PASS", "violations": []}, # fix me "V3": {"result": "PASS", "violations": []}, # fix me "V4": {"result": "PASS", "violations": []}, } # === end judgment === for vid, text in drafts: v = verdicts[vid] if v["result"] == "PASS": print(f"{vid}: PASS") else: print(f"{vid}: FAIL ({'; '.join(v['violations'])})") 31:T47c,PACK = { "banned": ["game-changing", "best-in-class", "revolutionary"], "max_words": 12, # subject-line cap — tighter than the 20-word body rule } drafts = [ ("V1", "Fresh beans at your door every Tuesday"), ("V2", "The game-changing subscription your mornings deserve"), ("V3", "We roast small batches weekly and ship them straight to your kitchen door for you"), ("V4", "Skip a week anytime, no questions"), ] def lint_copy(text, pack): problems = [] lowered = text.lower() for phrase in pack["banned"]: if phrase in lowered: problems.append(f"banned: {phrase}") if len(text.split()) > pack["max_words"]: problems.append("too long") return problems verdicts = { "V1": {"result": "PASS", "violations": []}, "V2": {"result": "FAIL", "violations": ["banned: game-changing"]}, "V3": {"result": "FAIL", "violations": ["too long"]}, "V4": {"result": "PASS", "violations": []}, } for vid, text in drafts: v = verdicts[vid] if v["result"] == "PASS": print(f"{vid}: PASS") else: print(f"{vid}: FAIL ({'; '.join(v['violations'])})") 32:T7de,PACK = { "banned": ["revolutionary", "best-in-class", "#1"], "max_words": 10, # Dana tightened the short-form cap from 12 after the first batch } variants = [ {"id": "E1", "channel": "email", "text": "Your Tuesday beans are on the way"}, {"id": "E2", "channel": "email", "text": "The revolutionary roast subscription"}, {"id": "S1", "channel": "social", "text": "Small batches, roasted this week, shipped this week"}, {"id": "S2", "channel": "social", "text": "The #1 best-in-class coffee experience for people who demand so much more"}, {"id": "A1", "channel": "ads", "text": "Skip a week anytime"}, ] # --- complete helpers (optional tools — do not rewrite) --- # Purpose: record what the pack would flag so you can compare to your HITL marks. # Gloss: `for phrase in pack["banned"]` means "check each banned phrase one at a time." def lint_copy(text, pack): problems = [] lowered = text.lower() for phrase in pack["banned"]: if phrase in lowered: problems.append(f"banned: {phrase}") if len(text.split()) > pack["max_words"]: problems.append("too long") return problems # === YOUR JUDGMENT — edit only this block === # call: "SHIP" (PASS / goes to Dana) or "HOLD" (FAIL / revision) # violations: labels when HOLD; empty when SHIP # Instruction to the model on HOLD (for your notes): drop banned X / cut to ≤N / both verdicts = { "E1": {"call": "SHIP", "violations": []}, "E2": {"call": "SHIP", "violations": []}, # fix me "S1": {"call": "SHIP", "violations": []}, "S2": {"call": "SHIP", "violations": []}, # fix me "A1": {"call": "SHIP", "violations": []}, } # === end judgment === kept = 0 for v in variants: d = verdicts[v["id"]] if d["call"] == "SHIP": print(f"{v['id']} {v['channel']:<6} PASS") kept += 1 else: print(f"{v['id']} {v['channel']:<6} FAIL {'; '.join(d['violations'])}") print(f"kept {kept}/{len(variants)} — batch goes to Dana for approval") 33:T5d5,PACK = { "banned": ["revolutionary", "best-in-class", "#1"], "max_words": 10, } variants = [ {"id": "E1", "channel": "email", "text": "Your Tuesday beans are on the way"}, {"id": "E2", "channel": "email", "text": "The revolutionary roast subscription"}, {"id": "S1", "channel": "social", "text": "Small batches, roasted this week, shipped this week"}, {"id": "S2", "channel": "social", "text": "The #1 best-in-class coffee experience for people who demand so much more"}, {"id": "A1", "channel": "ads", "text": "Skip a week anytime"}, ] def lint_copy(text, pack): problems = [] lowered = text.lower() for phrase in pack["banned"]: if phrase in lowered: problems.append(f"banned: {phrase}") if len(text.split()) > pack["max_words"]: problems.append("too long") return problems verdicts = { "E1": {"call": "SHIP", "violations": []}, "E2": {"call": "HOLD", "violations": ["banned: revolutionary"]}, "S1": {"call": "SHIP", "violations": []}, "S2": {"call": "HOLD", "violations": ["banned: best-in-class", "banned: #1", "too long"]}, "A1": {"call": "SHIP", "violations": []}, } kept = 0 for v in variants: d = verdicts[v["id"]] if d["call"] == "SHIP": print(f"{v['id']} {v['channel']:<6} PASS") kept += 1 else: print(f"{v['id']} {v['channel']:<6} FAIL {'; '.join(d['violations'])}") print(f"kept {kept}/{len(variants)} — batch goes to Dana for approval") 34:Tc7c,# Slots before prompts You're in the campaign studio. **What you're doing this lesson:** list every asset the Fieldnote fall relaunch needs — where it lives, how many versions, and the words on the image — before anyone opens an image tool. That list is the **batch plan**. You do not need the voice-pack lesson to start. The amateur move with image models is opening the tool and typing "cozy coffee vibes autumn." Forty regenerations later there's a folder called `final-final-v3` and nothing that fits the story format. The variable that was never pinned down: *where does this asset live?* Professionals plan the batch backwards from the **slots** (the destinations: email header, story frame, feed square, ad banner). The relaunch needs those four slots, four sizes, four different jobs. Write them as a list of records before you generate anything. Every record answers the questions that otherwise surface as rework: - **Slot and size** — the destination, named before generation starts. A tall story crop is not a wide banner with ambitions. Composition depends on the frame; the frame comes from the slot. - **Variants** — how many candidates to generate per slot. You generate several and *select*, because selection with a checklist beats regenerate-until-pretty on both cost and sanity. More variants for the ad banner (it carries the budget) than the email header (it's decorative). - **Text on the asset** — the exact words, or the explicit value "none." This field earns its keep in the next step, so clock it now: it's either precise copy — already checked against the brand voice — or a deliberate "no text." Fieldnote's plan for this relaunch, in English: six email-header candidates with no text on the image; eight story frames carrying "Fall tier. Same beans, better math."; eight feed squares with no text; ten ad banners carrying "Skip a week anytime." That's 32 generations, costed before anyone hits generate. ## Why the counts are in the plan Because counts are what turn "make some assets" into something you can cost and defend. At your image tool's per-image rate, 32 is a number — and a number means you can walk into the budget conversation with "the relaunch creative costs this much at this batch depth" instead of a shrug. When someone wants more ad variants, the plan shows exactly what the extra depth costs. You'll confirm that plan and the cut-first tradeoff two steps from now. The plan is also the review sheet's spine. When the batch comes back, every variant gets scored against its slot's brief — fits the frame, brand constraints met, text legible at target size, nothing weird in the corners. Same checks for every variant in the slot. A batch where everything passes review is a checklist that checks nothing; expect to reject, and keep the rejects scored so the next brief learns from them. One thing the plan deliberately does *not* contain: model names. Tools churn monthly; slots, sizes, variant counts, and text-on-asset are the stable layer. Write the plan against the work, point it at whatever generator you're using this quarter. First, though — a drill on the field people forget, and what forgetting it costs.35:T696,SLOTS = [ {"slot": "email-header", "variants": 6}, {"slot": "story", "variants": 8}, {"slot": "feed-square", "variants": 8}, {"slot": "ad-banner", "variants": 10}, ] COST_PER_IMAGE = 0.04 # --- complete helper (optional calculator — do not rewrite) --- # Purpose: total assets and dollars from a slot table so you can check your plan. # Gloss: `s["variants"] * cost_per_image` is count times rate for that line. def batch_plan(slots, cost_per_image): total_assets = 0 total_cost = 0.0 lines = [] for s in slots: line_cost = s["variants"] * cost_per_image lines.append((s["slot"], s["variants"], line_cost)) total_assets += s["variants"] total_cost += line_cost return lines, total_assets, total_cost # === YOUR JUDGMENT — edit only this block === # Mirror the required slots/counts; line_cost = variants * 0.04 plan = [ {"slot": "email-header", "variants": 0, "line_cost": 0.00}, # fix me {"slot": "story", "variants": 0, "line_cost": 0.00}, # fix me {"slot": "feed-square", "variants": 0, "line_cost": 0.00}, # fix me {"slot": "ad-banner", "variants": 0, "line_cost": 0.00}, # fix me ] total_assets = 0 # fix me — should be 32 total_cost = 0.00 # fix me — should be 1.28 # If generation budget halves, cut this slot first (decorative vs budget carrier) cut_first = "ad-banner" # fix me cut_why = "placeholder" # fix me — short reason # === end judgment === for row in plan: print(f"{row['slot']:<13} x{row['variants']:<3} ${row['line_cost']:.2f}") print(f"total: {total_assets} assets, ${total_cost:.2f}") print(f"cut-first: {cut_first} — {cut_why}") 36:T476,SLOTS = [ {"slot": "email-header", "variants": 6}, {"slot": "story", "variants": 8}, {"slot": "feed-square", "variants": 8}, {"slot": "ad-banner", "variants": 10}, ] COST_PER_IMAGE = 0.04 def batch_plan(slots, cost_per_image): total_assets = 0 total_cost = 0.0 lines = [] for s in slots: line_cost = s["variants"] * cost_per_image lines.append((s["slot"], s["variants"], line_cost)) total_assets += s["variants"] total_cost += line_cost return lines, total_assets, total_cost plan = [ {"slot": "email-header", "variants": 6, "line_cost": 0.24}, {"slot": "story", "variants": 8, "line_cost": 0.32}, {"slot": "feed-square", "variants": 8, "line_cost": 0.32}, {"slot": "ad-banner", "variants": 10, "line_cost": 0.40}, ] total_assets = 32 total_cost = 1.28 cut_first = "email-header" cut_why = "decorative; ad-banner carries the budget" for row in plan: print(f"{row['slot']:<13} x{row['variants']:<3} ${row['line_cost']:.2f}") print(f"total: {total_assets} assets, ${total_cost:.2f}") print(f"cut-first: {cut_first} — {cut_why}") 37:Ta36,# The calendar is data, and the approval column is the gate Ask AI to expand the relaunch brief into a two-week calendar and it does a genuinely good job: sensible cadence, channels alternating, launch-day cluster, a winback email at the end. Expanding a brief into a grid is exactly the kind of work models are built for. Then look at what it produced: a list of things that will be *published under your brand's name*, generated in eight seconds, reviewed by no one. The grid is not the dangerous part. The dangerous part is that every scheduling tool on earth will accept the grid as-is. So the calendar gets a schema, and the schema gets a gate: ``` date, channel, piece, owner, approved_by, status ``` Two of those columns do the safety work: - **`approved_by`** — a name, or empty. Not "marketing." Not "the team." A person who read the piece and put their name in the row. Scripts fill every other column; this one only ever gets filled by a human. When the FTC, a journalist, or your CEO asks who signed off on the post, this column is the answer — which is precisely why no script writes to it. - **`status`** — `approved` or `rejected`. And here's the counterintuitive rule: **rejected rows stay in the file.** Delete a rejected row and you've deleted the record of what the team said no to. Next month, when the model proposes the same meme-post again — it will — the rejected row is how you know it was already considered and killed, and why. Run the editor. The teaser story has `status: approved` but an empty `approved_by` — someone flipped the status in a hurry without putting a name on it. The gate holds it. Status is an opinion; a named approver is accountability. The gate demands both. ## The rule with teeth Only rows that pass the gate reach the scheduling tool. This is the chapter's spine applied to publishing: the AI proposes, your script filters, a named human approves, and *then* the scheduler gets its rows. Courses and vendors mostly treat generation and publication as one continuous motion — draft it, schedule it, done. The gate is you cutting that motion in half and putting a human in the seam. Two smaller rules ride along, both learned the embarrassing way. The timezone is stated once at the top of the file — a 9am send in the wrong timezone is a 3am send in your customers' inboxes. And no channel posts twice in one day without a stated reason; if the model stacks three stories on launch day, either write the reason into the row or reject two. Next: the gate runs on four rows, and you predict which pieces SCHEDULE before the tool confirms.38:T4b0,# Purpose: calendar publish queue with a named-approver policy switch. # Walkthrough: for each row, if approved AND (policy off OR name present) → queue. # Gloss: True means the named-approver rule is enforced. # Role: judgment payload = the policy constant; helper is complete. rows = [ {"date": "09-02", "channel": "email", "piece": "launch-email", "approved_by": "dana", "status": "approved"}, {"date": "09-02", "channel": "social", "piece": "teaser-story", "approved_by": "", "status": "approved"}, {"date": "09-03", "channel": "social", "piece": "meme-post", "approved_by": "dana", "status": "rejected"}, {"date": "09-05", "channel": "email", "piece": "winback", "approved_by": "sam", "status": "approved"}, ] REQUIRE_NAMED_APPROVER = ___ # True = hold nameless "approved" rows def publish_queue(rows): queue = [] for r in rows: if r["status"] != "approved": continue if REQUIRE_NAMED_APPROVER and not r["approved_by"]: continue queue.append(r) return queue q = publish_queue(rows) print(f"queue: {len(q)} rows") for r in q: print(f"{r['date']} {r['piece']} -> approved by {r['approved_by']}") 39:T85f,rows = [ {"date": "09-02", "channel": "email", "piece": "launch-email", "approved_by": "dana", "status": "approved"}, {"date": "09-02", "channel": "social", "piece": "teaser-story", "approved_by": "", "status": "approved"}, {"date": "09-03", "channel": "social", "piece": "meme-post", "approved_by": "dana", "status": "rejected"}, {"date": "09-04", "channel": "ads", "piece": "ad-set-a", "approved_by": "sam", "status": "approved"}, {"date": "09-05", "channel": "email", "piece": "winback", "approved_by": "sam", "status": "approved"}, ] # --- complete helper (optional tool — do not rewrite) --- # Purpose: same calendar gate a teammate might automate; compare to your triage. # Gloss: empty approved_by ("") means no named human signed the row. def gate(rows): result = {"schedule": [], "blocked": [], "rejected": []} for r in rows: if r["status"] == "rejected": result["rejected"].append(r) elif r["status"] == "approved" and r["approved_by"]: result["schedule"].append(r) else: result["blocked"].append(r) return result # === YOUR JUDGMENT — edit only this block === # route per piece: "SCHEDULE" | "BLOCKED" | "KEPT-REJECTED" verdicts = { "launch-email": "SCHEDULE", "teaser-story": "SCHEDULE", # fix me — nameless approver "meme-post": "SCHEDULE", # fix me "ad-set-a": "SCHEDULE", "winback": "SCHEDULE", } # === end judgment === schedule, blocked, rejected = [], [], [] for r in rows: route = verdicts[r["piece"]] if route == "SCHEDULE": schedule.append(r) elif route == "BLOCKED": blocked.append(r) else: rejected.append(r) for r in schedule: print(f"SCHEDULE {r['date']} {r['channel']:<6} {r['piece']} ({r['approved_by']})") for r in blocked: print(f"BLOCKED {r['date']} {r['channel']:<6} {r['piece']} (no approver)") for r in rejected: print(f"KEPT {r['date']} {r['channel']:<6} {r['piece']} (rejected, stays in the file)") print(f"{len(schedule)} scheduled, {len(blocked)} blocked, {len(rejected)} rejected-and-kept") 3a:T6d7,rows = [ {"date": "09-02", "channel": "email", "piece": "launch-email", "approved_by": "dana", "status": "approved"}, {"date": "09-02", "channel": "social", "piece": "teaser-story", "approved_by": "", "status": "approved"}, {"date": "09-03", "channel": "social", "piece": "meme-post", "approved_by": "dana", "status": "rejected"}, {"date": "09-04", "channel": "ads", "piece": "ad-set-a", "approved_by": "sam", "status": "approved"}, {"date": "09-05", "channel": "email", "piece": "winback", "approved_by": "sam", "status": "approved"}, ] def gate(rows): result = {"schedule": [], "blocked": [], "rejected": []} for r in rows: if r["status"] == "rejected": result["rejected"].append(r) elif r["status"] == "approved" and r["approved_by"]: result["schedule"].append(r) else: result["blocked"].append(r) return result verdicts = { "launch-email": "SCHEDULE", "teaser-story": "BLOCKED", "meme-post": "KEPT-REJECTED", "ad-set-a": "SCHEDULE", "winback": "SCHEDULE", } schedule, blocked, rejected = [], [], [] for r in rows: route = verdicts[r["piece"]] if route == "SCHEDULE": schedule.append(r) elif route == "BLOCKED": blocked.append(r) else: rejected.append(r) for r in schedule: print(f"SCHEDULE {r['date']} {r['channel']:<6} {r['piece']} ({r['approved_by']})") for r in blocked: print(f"BLOCKED {r['date']} {r['channel']:<6} {r['piece']} (no approver)") for r in rejected: print(f"KEPT {r['date']} {r['channel']:<6} {r['piece']} (rejected, stays in the file)") print(f"{len(schedule)} scheduled, {len(blocked)} blocked, {len(rejected)} rejected-and-kept") 3b:Td27,# A report is a draft over data You're in the campaign studio. **What you're doing this lesson:** check every number in the launch-week report against the **export** (the spreadsheet of actual counts the email platform downloaded). If you can't point at the cell — or show the math — the number doesn't ship. You do not need the earlier campaign lessons to start. Launch week is done. The relaunch email went to 18,200 people, and the platform export is sitting in your downloads folder. You paste it into the model and ask for a performance summary, because writing "open rate improved meaningfully week over week" by hand is nobody's favorite job and the model's narrative will be smoother than yours. The summary comes back in eight seconds and it reads great. That's the problem. It reads *equally great whether the numbers in it are real or not.* A language model producing a results paragraph is doing sentence-completion over your data, and when a rounder, more flattering number makes a better sentence, you sometimes get the better sentence. A claimed 3.6% when the export says 3.4%. An open count that appeared nowhere in the file. The failure is quiet, plausible, and wearing your formatting. So the studio rule: **every figure in an AI-drafted report is either traced to the export or it doesn't ship.** Traced means one of exactly two things: - **A lookup.** The report says 618 clicks; the export's clicks cell says 618. Match. - **A shown calculation.** The report says 3.4% **CTR** (click-through rate — clicks divided by a named base). You can point at 618 / 18,200. The formula is visible, the inputs are export cells. Anything else — however confident, however nicely phrased — is invented, and invented numbers in a campaign report are how teams make confident wrong decisions. Nobody re-checks a number that's already in a deck. The verification pass happens *before* the deck, or it never happens — which is why it runs as a scheduled check, not a heroic one-off: end of launch week, every campaign, pointed at that campaign's fresh export. Campaigns recur; so does the check. ## Watch the denominator **Open rate** is opens over *delivered*. CTR on this platform is clicks over *sent*. Same export, different **denominators** (the bottom of the fraction), and each printed line has to say which one it used. Worth knowing before you compare against a benchmark: most email platforms compute CTR over *delivered* (and click-to-open over opens), so an unlabeled "CTR" can mean three different numbers. This platform's export reports it over sent — the stated denominator is what keeps that from being a trap. A percentage that doesn't state its denominator isn't a fact yet — hold that thought, because the next read step shows the marketing industry doing this to itself with its own AI-adoption statistics. ## Why this is your check to run, not the model's You can ask the model to double-check its own summary, and you should — it catches some slips. But the model checking the model is the intern proofreading their own essay. The pass that counts is mechanical: walk every claim, compare it to the export, and mark TRACED or INVENTED with no opinion about narrative flow. You'll practice exactly that in the next two steps — first by predicting a CTR verdict, then by marking a full claim sheet TRACED/INVENTED.3c:T72a,export = {"sent": 18200, "delivered": 17654, "opens": 4914, "clicks": 618, "unsubs": 43} claims = [ {"label": "emails sent", "metric": "sent", "value": 18200}, {"label": "opens", "metric": "opens", "value": 4914}, {"label": "clicks", "metric": "clicks", "value": 655}, {"label": "unsubscribes", "metric": "unsubs", "value": 43}, ] # --- complete helper (optional receipt runner — do not rewrite) --- # Purpose: compare each claim value to the export cell; print TRACED/INVENTED. # Gloss: export[c["metric"]] looks up the real count for that metric name. def verify_numbers(claims, export): traced = 0 for c in claims: actual = export[c["metric"]] if c["value"] == actual: print(f"{c['label']:<13} {c['value']:>6} TRACED") traced += 1 else: print(f"{c['label']:<13} {c['value']:>6} INVENTED (export: {actual})") return traced # === YOUR JUDGMENT — edit only this block === # mark: "TRACED" or "INVENTED"; export_actual only needed when INVENTED receipts = { "emails sent": {"mark": "TRACED", "export_actual": None}, "opens": {"mark": "TRACED", "export_actual": None}, "clicks": {"mark": "TRACED", "export_actual": None}, # fix me "unsubscribes": {"mark": "TRACED", "export_actual": None}, } approval_ready = True # fix me — False unless every claim traced # === end judgment === traced = 0 for c in claims: r = receipts[c["label"]] if r["mark"] == "TRACED": print(f"{c['label']:<13} {c['value']:>6} TRACED") traced += 1 else: print(f"{c['label']:<13} {c['value']:>6} INVENTED (export: {r['export_actual']})") ready = "approval-ready" if approval_ready else "NOT approval-ready" print(f"{traced}/{len(claims)} traced — report is {ready}") 3d:T581,export = {"sent": 18200, "delivered": 17654, "opens": 4914, "clicks": 618, "unsubs": 43} claims = [ {"label": "emails sent", "metric": "sent", "value": 18200}, {"label": "opens", "metric": "opens", "value": 4914}, {"label": "clicks", "metric": "clicks", "value": 655}, {"label": "unsubscribes", "metric": "unsubs", "value": 43}, ] def verify_numbers(claims, export): traced = 0 for c in claims: actual = export[c["metric"]] if c["value"] == actual: print(f"{c['label']:<13} {c['value']:>6} TRACED") traced += 1 else: print(f"{c['label']:<13} {c['value']:>6} INVENTED (export: {actual})") return traced receipts = { "emails sent": {"mark": "TRACED", "export_actual": None}, "opens": {"mark": "TRACED", "export_actual": None}, "clicks": {"mark": "INVENTED", "export_actual": 618}, "unsubscribes": {"mark": "TRACED", "export_actual": None}, } approval_ready = False traced = 0 for c in claims: r = receipts[c["label"]] if r["mark"] == "TRACED": print(f"{c['label']:<13} {c['value']:>6} TRACED") traced += 1 else: print(f"{c['label']:<13} {c['value']:>6} INVENTED (export: {r['export_actual']})") ready = "approval-ready" if approval_ready else "NOT approval-ready" print(f"{traced}/{len(claims)} traced — report is {ready}") 3e:Tfb0,# Denominators, and the list with legal teeth Two things kill more marketing reports than bad math: percentages that don't say what they're a percentage of, and claims that were never allowed to exist in the first place. ## The denominator trap **Code role:** read-only demo of two true stats with different denominators — job literacy, not a coding drill. Run the editor. Both statistics are real, current, and about "AI in marketing" — and they differ by a factor of three because they count different things. Salesforce's own State of Marketing 2026 report headlines 75% *of marketers* having adopted AI: any use, by anyone, counts once. The CMO Survey's Spring 2026 wave reports AI in 24.2% *of marketing activities*, up from 13.1% — activity-weighted, so one marketer using AI for one task out of twenty barely moves it. A report that says "75% of marketing is AI now" has silently swapped denominators, and it's the kind of error that survives every read-through because the sentence *sounds* fine. The rule for your own reports is the numbers pass you just practiced, extended one level: every percentage states its denominator. "3.4% CTR (clicks / sent)." "27.8% open rate (opens / delivered)." A share of activities and a share of users are different facts; a report that can't tell them apart isn't a report, it's mood. ## Three lists and a fourth category Lesson 2's gate answered *who approves*. The brand-safety policy answers *what needs approving*. Three lists: - **AI may draft** — social posts, email subject lines, ad copy. Low blast radius, fully linted by lesson 1's pass, still gated by lesson 2's calendar before publishing. - **Named approver required** — pricing claims, press quotes, partner mentions, anything with a competitor's name in it. Draftable, but a specific human signs before it moves. - **Human-only** — crisis comms, apologies, anything responding to a customer's bad day. The model doesn't draft these because the cost of sounding synthetic exceeds the cost of writing it yourself. And then a fourth category that's not a list, because it's not a choice: **banned outright.** The FTC's Consumer Reviews and Testimonials Rule — 16 CFR Part 465, effective October 21, 2024 — prohibits reviews and testimonials "by someone who does not exist, such as AI-generated fake reviews," with civil penalties in the tens of thousands of dollars per knowing violation (the figure is inflation-adjusted every year — look up the current year's amount before you ever quote a specific number). Note the unit: *per violation*. A batch of twelve invented five-star reviews is not one mistake. An AI-drafted testimonial from a customer the model made up goes on no list — not draftable, not approvable-with-sign-off, gone. And more broadly, the FTC's deceptive-practices standard doesn't care who wrote the sentence; "the AI drafted it" has the same legal weight as "the intern drafted it," which is none. ## Where these checks live You now hold three working checks, and none of them belongs in a notebook you remember to open. Each has a slot in a real marketing stack: - **Voice lint** runs where drafts are born: between "the model returns variants" and "variants land in Dana's approval queue." Every batch walks through it, watched or not. - **Calendar gate** is the pre-send doorway: rows need approved status and a named approver or they never reach the scheduler. - **Numbers pass** is scheduled end of launch week: every campaign, pointed at that campaign's export, before the report leaves the team. No vendor names in that list, for the same reason the batch plan carries no model names: tools churn quarterly, but the stack *roles* — draft check, pre-send gate, report receipt — are stable. A check without a slot quietly stops running by October. The next two steps drill both halves of the policy layer: first you'll spot which drafted line can't ship at any approval level, then you'll set the default tier for pieces that appear on no list.3f:Ta07,export = {"sent": 18200, "opens": 4914, "clicks": 618, "signups": 212} report_claims = [ {"metric": "sent", "value": 18200}, {"metric": "opens", "value": 4914}, {"metric": "clicks", "value": 618}, {"metric": "signups", "value": 230}, ] queue = [ {"piece": "results-recap email", "approved_by": "dana"}, {"piece": "we-hit-record post", "approved_by": ""}, ] # --- complete helper (optional observability — do not rewrite) --- # Purpose: show what an automated dual-gate audit would print; your # findings dict is still what the harness uses to pass. # Gloss: non-empty approved_by means a named human signed that piece. def campaign_audit(claims, export, queue): invented = 0 for c in claims: actual = export[c["metric"]] if c["value"] == actual: print(f"{c['metric']:<8} {c['value']:>6} TRACED") else: print(f"{c['metric']:<8} {c['value']:>6} INVENTED (export: {actual})") invented += 1 blocked = 0 for r in queue: if r["approved_by"]: print(f"{r['piece']}: OK ({r['approved_by']})") else: print(f"{r['piece']}: BLOCKED (no approver)") blocked += 1 if invented == 0 and blocked == 0: print("verdict: CLEARED FOR PUBLISH") else: print(f"verdict: HELD ({invented} invented number(s), {blocked} unapproved row(s))") # === YOUR JUDGMENT — edit only this block === number_marks = { "sent": {"mark": "TRACED", "export_actual": None}, "opens": {"mark": "TRACED", "export_actual": None}, "clicks": {"mark": "TRACED", "export_actual": None}, "signups": {"mark": "TRACED", "export_actual": None}, # fix me } approver_marks = { "results-recap email": {"lane": "OK", "name": "dana"}, "we-hit-record post": {"lane": "OK", "name": "dana"}, # fix me } # "CLEARED FOR PUBLISH" or "HELD (1 invented number(s), 1 unapproved row(s))" publish_verdict = "CLEARED FOR PUBLISH" # fix me # === end judgment === invented = 0 for c in report_claims: m = number_marks[c["metric"]] if m["mark"] == "TRACED": print(f"{c['metric']:<8} {c['value']:>6} TRACED") else: print(f"{c['metric']:<8} {c['value']:>6} INVENTED (export: {m['export_actual']})") invented += 1 blocked = 0 for r in queue: a = approver_marks[r["piece"]] if a["lane"] == "OK": print(f"{r['piece']}: OK ({a['name']})") else: print(f"{r['piece']}: BLOCKED (no approver)") blocked += 1 print(f"verdict: {publish_verdict}") 40:T868,export = {"sent": 18200, "opens": 4914, "clicks": 618, "signups": 212} report_claims = [ {"metric": "sent", "value": 18200}, {"metric": "opens", "value": 4914}, {"metric": "clicks", "value": 618}, {"metric": "signups", "value": 230}, ] queue = [ {"piece": "results-recap email", "approved_by": "dana"}, {"piece": "we-hit-record post", "approved_by": ""}, ] def campaign_audit(claims, export, queue): invented = 0 for c in claims: actual = export[c["metric"]] if c["value"] == actual: print(f"{c['metric']:<8} {c['value']:>6} TRACED") else: print(f"{c['metric']:<8} {c['value']:>6} INVENTED (export: {actual})") invented += 1 blocked = 0 for r in queue: if r["approved_by"]: print(f"{r['piece']}: OK ({r['approved_by']})") else: print(f"{r['piece']}: BLOCKED (no approver)") blocked += 1 if invented == 0 and blocked == 0: print("verdict: CLEARED FOR PUBLISH") else: print(f"verdict: HELD ({invented} invented number(s), {blocked} unapproved row(s))") number_marks = { "sent": {"mark": "TRACED", "export_actual": None}, "opens": {"mark": "TRACED", "export_actual": None}, "clicks": {"mark": "TRACED", "export_actual": None}, "signups": {"mark": "INVENTED", "export_actual": 212}, } approver_marks = { "results-recap email": {"lane": "OK", "name": "dana"}, "we-hit-record post": {"lane": "BLOCKED", "name": ""}, } publish_verdict = "HELD (1 invented number(s), 1 unapproved row(s))" invented = 0 for c in report_claims: m = number_marks[c["metric"]] if m["mark"] == "TRACED": print(f"{c['metric']:<8} {c['value']:>6} TRACED") else: print(f"{c['metric']:<8} {c['value']:>6} INVENTED (export: {m['export_actual']})") invented += 1 blocked = 0 for r in queue: a = approver_marks[r["piece"]] if a["lane"] == "OK": print(f"{r['piece']}: OK ({a['name']})") else: print(f"{r['piece']}: BLOCKED (no approver)") blocked += 1 print(f"verdict: {publish_verdict}") 41:T528,# When a number looks like news, take it to your analyst The numbers pass catches figures that don't match the export. A different claim gets straight through it: a number that matches the export exactly and still points the wrong way. Ask the model what else the export shows and it will find patterns. Some of them are real leads, and running that first pass yourself is a good use of AI. You get a lead in minutes instead of waiting a week for someone to have time. The catch is that a pattern in one export comes with a guess about why, and testing the why is someone else's craft. Your analyst, or whoever owns the data (the email platform admin, the finance team), knows whether the groups being compared are alike and what else changed that week. So before a pattern changes anything, send them a short note with four parts: 1. **The question.** What decision is waiting on this? 2. **The file.** Which export you used, and which rows or tab. 3. **The claim.** What the AI said, with its numbers. 4. **What would change your mind.** The result that would make you drop the idea. A note like that takes five minutes and gets a real answer. "AI says morning sends win, can you confirm?" gets a quick yes from someone busy, or no reply at all. One more call on the relaunch export, and then the lesson is done.42:T7f5,# One recording, two weeks of clips You're in the campaign studio. **What you're doing this lesson:** turn one long recording into a short list of clips you'd put Fieldnote's name on, then give each clip a day on the calendar. There is no code in this lesson, and every graded step is a call you'd make at work. Fieldnote ran a live tasting webinar last Thursday to launch the fall tier. Marta Lind, the head roaster, brewed four coffees and took questions from viewers. The recording sits in a shared folder, and very few people will ever watch the whole thing. Short vertical clips are how most of the relaunch audience will meet Marta, so your job is to find the best moments, check them, and schedule them. Here's what you have, before you start: - **The recording:** 48 minutes, with a timestamped transcript of about 7,400 words. - **The audience so far:** 520 people registered and 214 watched live. - **The slots:** five short clips across the two-week push, on days 1, 3, 6, 9 and 12. - **Dana's limits:** every clip runs under 60 seconds, is vertical, and has captions. - **The price sheet:** the fall tier is $16 for a 12 oz bag. The same bag costs $19.50 in the Fieldnote shop, so the tier saves $3.50 a bag. - **The shipping promise in the brief:** beans ship the Monday after they're roasted, and subscribers can skip a week anytime. - **The approvers:** Dana or Sam signs every row, the same as the calendar in lesson 2. **What you'll make:** a **clip plan**. It's one table listing the clips you'll post, where each one starts and ends, the line of text on screen, the day it goes out and who approved it. Under the table you list the clips you cut and why. **How the lesson runs:** 1. Read the one-paragraph prompt Fieldnote used and the eight moments the AI found. 2. Pick the clip that opens the push. 3. Learn the four checks every clip has to pass. 4. Make three calls: a wrong price, a line of screen text, and a plan row that isn't ready. 5. Take the clip plan template with you. It takes about 22 minutes.43:Ta15,# The prompt, and the eight moments it found Fieldnote pasted the transcript into its AI assistant with this one paragraph: > Here is the timestamped transcript of Fieldnote's 48-minute fall tasting webinar with our head roaster, Marta Lind. Find eight moments we could post as short vertical clips for the fall tier launch. Each clip has to make sense to someone who never saw the webinar, open on a line that makes a coffee drinker keep watching, and run 20 to 60 seconds. For each one, give the start and end time, the exact opening line, one sentence on why it works, and a flag for anything that needs checking: prices, numbers, health claims, and any viewer's name read aloud. Score each one from 1 to 10 for how likely people are to keep watching. Quote Marta exactly and never reword what anyone said. The prompt does three jobs. It states the rules a clip has to meet, it asks for exact words so you can check them against the recording, and it asks the AI to flag risky lines instead of quietly fixing them. Here's what came back: | # | Time | Length | Opening line | AI score | AI flag | |---|---|---|---|---|---| | C1 | 06:12 to 06:51 | 39 s | "Most people ruin good coffee in the last ten seconds." | 8 | none | | C2 | 11:40 to 12:35 | 55 s | "The fall tier saves you about six dollars a bag over buying the same beans in our shop." | 9 | price | | C3 | 17:05 to 17:38 | 33 s | "So like I said before, that's why the second cup tastes sweeter." | 7 | none | | C4 | 22:10 to 23:02 | 52 s | "Thanks, Priya from Denver. Great question about decaf." | 7 | viewer name | | C5 | 28:44 to 29:20 | 36 s | "Coffee is a fruit, and this cup tastes like it." | 8 | none | | C6 | 34:00 to 35:30 | 90 s | "The first time I roasted this bean, I burned the whole batch." | 9 | over 60 s | | C7 | 39:15 to 39:48 | 33 s | "This one keeps me sharp all afternoon, with none of the jitters." | 6 | health claim | | C8 | 44:30 to 45:05 | 35 s | "Your beans ship the Monday after we roast them, and you can skip a week anytime." | 6 | none | The score column is the AI's guess at how long people will watch. It tells you nothing about whether a clip is true, fits the slot, or makes sense on its own, and the two highest scores both have problems you'll find in the next steps. Run every clip through the checks yourself and use the score to break ties. The AI got you from 48 minutes to eight candidates in under a minute. Scrubbing the recording by hand would have taken most of an afternoon. That time saved is the reason to use it, and the checks are where your judgment comes in.44:T7c7,# Four checks before a clip gets a slot The AI handles the first pass well. It reads 7,400 words in seconds and finds moments a person would spend an afternoon looking for. What it can't judge is what Fieldnote should post under its own name, so these four checks are yours. **1. It makes sense alone.** Someone scrolling has never heard of Marta or the webinar. The first line has to work with no setup, and the clip has to end where a thought ends. Openings like "like I said" or "going back to that" mean the clip starts in the wrong place. Moving the start a few seconds often fixes it, so check the transcript before you throw a candidate out. **2. It fits the slot.** Dana's limit is 60 seconds. When a moment runs long, trim the start and end down to its strongest part. Speeding the audio up to fit makes Marta sound rushed, so trim instead. **3. Every word is true and approved.** Any price, number or promise in a clip gets traced to a source, the same way the numbers pass traced the report to the export. Captions must match what Marta says word for word, which means a wrong number can't be fixed in the caption, because the audio still says it. AI captions also mishear names. Expect "Fieldnote" to come back as "field note" and "Swiss Water decaf" as "sweet water decaf", and read every caption before it posts. Claims about health or focus stay out entirely, because the fall tier brief only allows claims about taste, price and shipping. **4. Everyone in it agreed.** Marta agreed to have clips cut from her webinar. Viewers agreed to a recorded session, and that's a smaller thing than appearing in a promoted clip. When a viewer's name is read aloud, start the clip after the name. The line of text you put on screen goes through the same checks. The AI will suggest one for each clip, and that line is copy like any other: it passes the voice pack from lesson 1 (no banned phrases, 10 words or fewer for short-form) and it promises nothing the clip doesn't deliver.45:T487,# The posting plan Each clip that passes the checks becomes a row on the calendar from lesson 2. The gate works the same way here: a row without a named approver never reaches the scheduler. Every row needs these columns: | Column | What goes in it | |---|---| | Day | The day of the push it goes out | | Clip | The candidate number | | Cut | Start and end time, after any trimming | | Screen line | The on-screen text, 10 words or fewer | | Channels | Where it posts | | Captions checked by | The person who read every caption against the audio | | Approved by | Dana or Sam | | Links to | The full recording or the fall tier page | Two of these columns get skipped most often. **Captions checked by** names a person, because the tool that wrote the captions can't catch its own mishearing. **Links to** gives every clip somewhere to send people, since a clip that gets views and sends nobody to the fall tier page hasn't helped the launch. When the push ends, the clip results go through the numbers pass with the rest of the campaign. Views, watch time and clicks come from each platform's own export, never from a screenshot someone dropped in Slack.46:T5c7,# Your clip plan The deliverable from this lesson is the **clip plan**: the posting table, plus a short list of the clips you cut and the reason for each. Here's Fieldnote's, with C4 moved to start on Marta's answer: | Day | Clip | Cut | Screen line | Approved by | |---|---|---|---|---| | 1 | C5 | 28:44 to 29:20 | This cup tastes like the fruit it came from | Dana | | 3 | C1 | 06:12 to 06:51 | The last ten seconds of brewing matter most | Sam | | 6 | C8 | 44:30 to 45:05 | Fresh roast every week. Skip a week anytime | Dana | | 9 | C4 | 22:16 to 23:02 | Decaf that still tastes like coffee | Dana | | 12 | C6 | 34:22 to 35:07 | The batch Marta burned, and what it taught her | Sam | Cut, with reasons: - **C2:** Marta says about $6 a bag and the price sheet says $3.50. Held, and Marta has the right number for her next webinar. - **C3:** it opens mid-thought, and the setup it depends on is more than a minute earlier in the recording. - **C7:** a focus and jitters claim, which the fall tier brief doesn't allow. To run this on your own recording, keep the prompt from step 2 and change four things: who is speaking, what the clips are for, your length limit, and what counts as a risky line in your business. For a software company that list could be prices, customer results and client names. Then run the same four checks and fill in the same table. The AI saves you an afternoon of scrubbing through the recording, and the four checks are where your judgment goes.28:["$","$L2b",null,{"tree":{"toc":{"version":"0.1.0","generatedAt":"1970-01-01T00:00:00.000Z","chapters":[{"number":0,"slug":"before-you-build","title":"from worker to builder","blurb":"stop doing repeated work by hand. this chapter is the map: the builder loop, old workplace deliverables turned into site artifacts, and the reusable tool card. if you are starting here, those moves are the lesson.{{#path.not:marketer}} if you already ran a studio, you already used the loop.{{/path.not}}{{#path.is:marketer}} the labs come next. campaign studio is {{path.ref:campaign-studio}}.{{/path.is}}","status":"live","lessonCount":11,"stepCount":22,"liveLessonCount":11,"liveStepCount":22,"estMinutes":114,"xpTotal":55},{"number":1,"slug":"read-the-script","title":"read the script — your first code-reading lab","blurb":"ai hands you small scripts. this lab covers the three reads that matter: what the names hold, whether the function actually returns, and which branch of the if-chain really fires. if you are starting here, this labels the reads. if you already ran a studio, you already wrote these.","status":"live","lessonCount":3,"stepCount":24,"liveLessonCount":3,"liveStepCount":24,"estMinutes":21,"xpTotal":78},{"number":2,"slug":"read-the-data-shapes","title":"read the data shapes — the json ai just dumped","blurb":"every api response, config file, and llm reply is lists and dicts nested in a tree. read the shape on sight, walk it with a loop, and catch the wrong-layer bug ai loves to ship.","status":"live","lessonCount":3,"stepCount":24,"liveLessonCount":3,"liveStepCount":24,"estMinutes":23,"xpTotal":80},{"number":3,"slug":"read-the-failure","title":"read the failure — cursor wrote this and crashed","blurb":"when python crashes, it tells you exactly what happened and where. most non-engineers panic at the wall of text — and ai 'fixes' the crash by hiding it. learn to read the failure, then catch only what's worth catching.","status":"live","lessonCount":3,"stepCount":24,"liveLessonCount":3,"liveStepCount":24,"estMinutes":24,"xpTotal":83},{"number":4,"slug":"read-the-pipeline","title":"read the pipeline — files in, api calls out","blurb":"every real script reads a file or calls an api — usually both. learn the with-block, the status-code families, and the retry rules: the seatbelts ai forgets to put on.","status":"live","lessonCount":3,"stepCount":24,"liveLessonCount":3,"liveStepCount":24,"estMinutes":26,"xpTotal":83},{"number":7,"slug":"mutation-and-state","title":"mutation — why your code mysteriously breaks","blurb":"when a list inside a function changes the list outside the function, that's mutation. ai does this constantly without flagging it, and it's the bug class that takes the longest to find.","status":"live","lessonCount":2,"stepCount":9,"liveLessonCount":2,"liveStepCount":9,"estMinutes":10,"xpTotal":32},{"number":13,"slug":"llm-apis","title":"llm apis — talking to claude and openai","blurb":"every ai feature you ship eventually calls a model api. learn the messages pattern, how to read the response, and the four lines ai writes every single time.","status":"live","lessonCount":4,"stepCount":27,"liveLessonCount":4,"liveStepCount":27,"estMinutes":45,"xpTotal":87},{"number":14,"slug":"structured-output","title":"structured output — making the model return json","blurb":"free-form text breaks every pipeline. learn the schema-first pattern ai uses to get reliable json back, validate it with pydantic, and catch the model's lies before they hit prod.","status":"live","lessonCount":3,"stepCount":25,"liveLessonCount":3,"liveStepCount":25,"estMinutes":38,"xpTotal":80},{"number":15,"slug":"mcp","title":"mcp — the model context protocol","blurb":"mcp is the new standard for plugging tools and data sources into ai agents. learn what an mcp server actually is, how claude code lists tools, and why this is replacing one-off integrations everywhere.","status":"live","lessonCount":3,"stepCount":26,"liveLessonCount":3,"liveStepCount":26,"estMinutes":40,"xpTotal":85},{"number":16,"slug":"agent-loops","title":"agent loops — tool_use and the request/tool/respond cycle","blurb":"an agent isn't magic. it's a while loop. learn the actual cycle claude code, cursor, and every other agent uses: model returns tool_use, you run the tool, you send the result back, repeat until end_turn.","status":"live","lessonCount":6,"stepCount":48,"liveLessonCount":6,"liveStepCount":48,"estMinutes":71,"xpTotal":154},{"number":17,"slug":"git-and-github","title":"git and github cli — the ai builder's actual workflow","blurb":"cursor and claude code commit on your behalf. reading those commits — and undoing the bad ones — is your job. learn the three-state model, the commands you'll run every day, and what `gh` does that `git` can't.","status":"live","lessonCount":3,"stepCount":24,"liveLessonCount":3,"liveStepCount":24,"estMinutes":36,"xpTotal":79},{"number":18,"slug":"secrets-and-env","title":"secrets — .env, api keys, and what not to commit","blurb":"ai ships keys to github all the time. learn the .env pattern, why os.getenv is non-negotiable, what to do when a key leaks, and the gitignore lines you need on day one.","status":"live","lessonCount":2,"stepCount":16,"liveLessonCount":2,"liveStepCount":16,"estMinutes":16,"xpTotal":58},{"number":19,"slug":"prompting","title":"prompting ai coding tools","blurb":"an AI coding tool writes and edits with you — Cursor is one brand, Claude Code is another. the difference between a one-shot session and a four-hour spiral is almost always the first prompt.","status":"live","lessonCount":5,"stepCount":36,"liveLessonCount":5,"liveStepCount":36,"estMinutes":59,"xpTotal":118},{"number":20,"slug":"agent-traces","title":"reading agent traces and telemetry","blurb":"when an agent fails, the trace tells you exactly where. learn to read tool calls, tool results, and stop reasons — the json breadcrumbs every agent leaves behind.","status":"live","lessonCount":3,"stepCount":25,"liveLessonCount":3,"liveStepCount":25,"estMinutes":30,"xpTotal":88},{"number":21,"slug":"evals","title":"test the agent before you trust it","blurb":"same input, check the output, pass or fail. don't ship on a demo that felt right.","status":"live","lessonCount":5,"stepCount":34,"liveLessonCount":5,"liveStepCount":34,"estMinutes":57,"xpTotal":112},{"number":22,"slug":"context-and-retrieval","title":"context and retrieval: feeding the model real data","blurb":"rag without the overengineering. chunking, embeddings, vector search, and the small set of patterns that make a model answer from your data instead of its training set.","status":"live","lessonCount":6,"stepCount":48,"liveLessonCount":6,"liveStepCount":48,"estMinutes":85,"xpTotal":150},{"number":23,"slug":"production-tradeoffs","title":"production tradeoffs — cost, latency, quality","blurb":"the three numbers every shipped llm feature lives or dies by. token math, caching, streaming, batching, and the small set of decisions that move the product more than a model swap ever will.","status":"live","lessonCount":4,"stepCount":33,"liveLessonCount":4,"liveStepCount":33,"estMinutes":57,"xpTotal":111},{"number":24,"slug":"debugging-output","title":"debugging broken ai output","blurb":"{{#path.is:copywriter}}when the model ships a claim nobody sourced.{{/path.is}}{{#path.is:lawyer}}when the model ships a privilege call nobody cited.{{/path.is}}{{#path.not:copywriter}}{{#path.not:lawyer}}when the model lies to your customer.{{/path.not}}{{/path.not}} the methodology for narrowing down what went wrong, the four most-common breakage classes, and the discipline that separates 'we shipped a fix' from 'we blamed the model and shrugged'.","status":"live","lessonCount":4,"stepCount":34,"liveLessonCount":4,"liveStepCount":34,"estMinutes":61,"xpTotal":112},{"number":25,"slug":"capstone","title":"path checkpoint — ship the CLI agent","blurb":"wire it all together. the prompt, the call, the validation, the trace, the eval, the MCP tool. a path checkpoint: the smallest end-to-end llm feature you could ship to a real user. on the developer route, harnesses, image, video, and production tooling still follow. retrieval and prompt-cache cost work live in later chapters — extend this build with them when you scale past the demo input set.","status":"live","lessonCount":7,"stepCount":62,"liveLessonCount":7,"liveStepCount":62,"estMinutes":100,"xpTotal":219},{"number":26,"slug":"agent-harnesses","title":"agent harnesses — the layer between you and the raw api","blurb":"claude code, cursor, aider, codex cli — they're all the same four layers wrapped around the same model api. learn what those layers are, what each adds, and what you'd build yourself if you had to. then you build a small harness you can run and break. when the harness changes the world, the rollback lives in the repo — a runbook a junior can execute at 3am.","status":"live","lessonCount":6,"stepCount":46,"liveLessonCount":6,"liveStepCount":46,"estMinutes":87,"xpTotal":151},{"number":27,"slug":"ai-image-generation","title":"ai image generation — from prompt to production asset","blurb":"the 2026 image model landscape, the prompts that work, and the pipeline that turns one good idea into a hundred ready-to-ship images. nano banana 2 for volume, nano banana pro for fidelity, flux, midjourney, ideogram, gpt-image-2. when each wins and what they cost.","status":"live","lessonCount":3,"stepCount":24,"liveLessonCount":3,"liveStepCount":24,"estMinutes":60,"xpTotal":67},{"number":28,"slug":"ai-video-generation","title":"ai video generation — omni, veo 3.1, sora, and the second wave","blurb":"video is the hardest content type to generate, the most expensive, and the most strategically interesting. learn the 2026 model lineup, the camera-control patterns that separate slop from craft, and the cost math that decides whether your idea is viable.","status":"live","lessonCount":3,"stepCount":23,"liveLessonCount":3,"liveStepCount":23,"estMinutes":50,"xpTotal":61},{"number":29,"slug":"programmatic-design","title":"programmatic design — code-driven video and the new design pipeline","blurb":"ai generates raw assets; code stitches them into something shippable. hyperframes, remotion, claude design — when each tool wins, how they combine, and the data-driven workflows that turn one template into a hundred videos.","status":"live","lessonCount":3,"stepCount":24,"liveLessonCount":3,"liveStepCount":24,"estMinutes":46,"xpTotal":66},{"number":30,"slug":"harness-engineering","title":"harness engineering — the discipline behind the model","blurb":"every coding agent is a model plus a harness. the model is bought; the harness is engineered. learn the craft: how to ratchet rules from failures, fight context rot, design long-horizon loops, wire hooks as enforcement, and read the haas shift that's reshaping what you build vs buy.","status":"live","lessonCount":6,"stepCount":52,"liveLessonCount":6,"liveStepCount":52,"estMinutes":126,"xpTotal":143},{"number":31,"slug":"intro-to-terminal","title":"intro to terminal — the text way to drive your computer","blurb":"the terminal is a text box: type a name, press enter, the computer does it. by the end you can move around your files, make folders, and read files without the mouse. every later tool chapter assumes this.","status":"live","lessonCount":3,"stepCount":23,"liveLessonCount":3,"liveStepCount":23,"estMinutes":26,"xpTotal":50},{"number":32,"slug":"intro-to-claude-cli","title":"intro to claude cli — the door to real building","blurb":"you've used claude in a chat window. the claude cli is the same model with its hands on your actual files. this chapter installs it, signs you in, and runs your first real command. by the end you've watched an ai read, plan, and change things on your machine, and you know when to reach for the cli instead of the chat box.","status":"live","lessonCount":3,"stepCount":23,"liveLessonCount":3,"liveStepCount":23,"estMinutes":32,"xpTotal":48},{"number":33,"slug":"intro-to-codex-cli","title":"intro to openai codex cli — a second tool in the kit","blurb":"you know the claude cli. the openai codex cli does the same job, an ai working in your terminal on your real files, with a different company behind it. this chapter installs it, signs you in, and runs your first command. most of what you already know carries straight over, so this chapter is mostly about what is different and when to reach for which.","status":"live","lessonCount":3,"stepCount":23,"liveLessonCount":3,"liveStepCount":23,"estMinutes":32,"xpTotal":45},{"number":34,"slug":"team-skills","title":"claude skills for teams — playbooks your whole team shares","blurb":"a claude skill is a packaged set of instructions — your team's playbook — that claude loads when it's relevant, so nobody has to re-explain it. this chapter is for people who manage teams. it covers what a skill is, how a team shares and provisions skills, real examples for hr, legal, and ops work, when a skill beats a one-off prompt, and the governance you need before any skill touches real work.","status":"live","lessonCount":4,"stepCount":28,"liveLessonCount":4,"liveStepCount":28,"estMinutes":41,"xpTotal":60},{"number":35,"slug":"dataframes-numpy-pandas","title":"dataframes with numpy and pandas","blurb":"tables are the working surface of applied ml. learn rows, columns, missing values, joins, aggregates, and the dataframe habits ai-generated notebooks assume.","status":"live","lessonCount":6,"stepCount":42,"liveLessonCount":6,"liveStepCount":42,"estMinutes":43,"xpTotal":156},{"number":36,"slug":"sql-for-ml-datasets","title":"sql for ml datasets","blurb":"most training data starts in a database. learn the select, join, filter, aggregate, and leakage traps that decide whether a model is learning signal or nonsense.","status":"live","lessonCount":4,"stepCount":28,"liveLessonCount":4,"liveStepCount":28,"estMinutes":28,"xpTotal":104},{"number":37,"slug":"dataset-pipelines","title":"dataset formats, ingestion, and validation pipelines","blurb":"a dataset is a product surface. build ingestion, validation, partitions, manifests, and checkpoints so the next run is not a mystery.","status":"live","lessonCount":5,"stepCount":35,"liveLessonCount":5,"liveStepCount":35,"estMinutes":37,"xpTotal":130},{"number":38,"slug":"ml-math-and-stats","title":"ml math and statistics that actually show up","blurb":"vectors, probability, distributions, correlation, and uncertainty are not trivia. they are how you read model behavior without worshipping it.","status":"live","lessonCount":5,"stepCount":35,"liveLessonCount":5,"liveStepCount":35,"estMinutes":37,"xpTotal":130},{"number":39,"slug":"supervised-learning-workflows","title":"supervised learning workflows","blurb":"labels, splits, baselines, training, prediction, and evaluation. the supervised workflow is the first complete model loop.","status":"live","lessonCount":5,"stepCount":35,"liveLessonCount":5,"liveStepCount":35,"estMinutes":40,"xpTotal":130},{"number":40,"slug":"unsupervised-learning-and-embeddings","title":"unsupervised learning, embeddings, and recommenders","blurb":"not every useful model has labels. cluster, compare, retrieve, and recommend by turning examples into useful neighborhoods.","status":"live","lessonCount":4,"stepCount":28,"liveLessonCount":4,"liveStepCount":28,"estMinutes":32,"xpTotal":104},{"number":41,"slug":"metrics-and-error-analysis","title":"metrics, slices, and error analysis","blurb":"accuracy is a blunt instrument. learn confusion matrices, precision, recall, thresholds, slices, and failure notes so model quality has evidence.","status":"live","lessonCount":5,"stepCount":35,"liveLessonCount":5,"liveStepCount":35,"estMinutes":37,"xpTotal":130},{"number":42,"slug":"pytorch-tensors-and-autograd","title":"pytorch tensors and autograd","blurb":"read tensor code without flinching. tensors, shapes, broadcasting, gradients, and autograd are the grammar of modern deep learning scripts.","status":"live","lessonCount":4,"stepCount":28,"liveLessonCount":4,"liveStepCount":28,"estMinutes":30,"xpTotal":104},{"number":43,"slug":"training-loops-and-optimizers","title":"training loops, backprop, optimizers, and schedulers","blurb":"the training loop is where models change. learn loss, gradients, optimizer steps, schedules, checkpoints, and the bugs ai ships there.","status":"live","lessonCount":5,"stepCount":35,"liveLessonCount":5,"liveStepCount":35,"estMinutes":37,"xpTotal":130},{"number":44,"slug":"deep-learning-architectures","title":"cnns, transformers, and useful llm internals","blurb":"architecture literacy for builders: convolution, attention, tokens, decoding, kv cache, quantization, and what those choices do to cost and behavior.","status":"live","lessonCount":5,"stepCount":35,"liveLessonCount":5,"liveStepCount":35,"estMinutes":37,"xpTotal":130},{"number":45,"slug":"feature-experiments-registries","title":"feature pipelines, experiment tracking, and registries","blurb":"features, runs, configs, artifacts, and registries are how ml work becomes repeatable instead of a lucky notebook.","status":"live","lessonCount":5,"stepCount":35,"liveLessonCount":5,"liveStepCount":35,"estMinutes":37,"xpTotal":130},{"number":46,"slug":"model-serving-and-mlops","title":"model serving, ci/cd, and mlops","blurb":"serving a model means handling inputs, versions, routes, batch jobs, ci gates, rollback, and production failures deliberately.","status":"live","lessonCount":6,"stepCount":42,"liveLessonCount":6,"liveStepCount":42,"estMinutes":44,"xpTotal":156},{"number":47,"slug":"monitoring-cloud-portfolio","title":"monitoring, drift, cloud scale, and portfolio launch","blurb":"the last mile: logs, drift, alerts, retraining decisions, cloud cost, gpu constraints, architecture docs, demos, and role stories.","status":"live","lessonCount":6,"stepCount":42,"liveLessonCount":6,"liveStepCount":42,"estMinutes":44,"xpTotal":156},{"number":48,"slug":"graphic-design-studio","title":"graphic design studio — prompt, kit, and fifty on-brand assets","blurb":"how a working designer runs image models: prompts written like creative briefs, a brand bible that travels as structured data — with the design tokens that carry it onto screens — and the seeds, reference locks, and QA filters that make fifty generated assets look like one studio shipped them.","status":"live","lessonCount":3,"stepCount":28,"liveLessonCount":3,"liveStepCount":28,"estMinutes":75,"xpTotal":84},{"number":49,"slug":"campaign-studio","title":"campaign studio: one campaign, end to end, with the gate on","blurb":"Run the Fieldnote fall relaunch end to end: copy checked against the voice pack, a batch plan, a calendar where every post has a named approver, a results report where every number traces to the export, and a clip plan that turns one webinar into short clips you can post. Built for the person who owns the send button.","status":"live","lessonCount":4,"stepCount":33,"liveLessonCount":4,"liveStepCount":33,"estMinutes":85,"xpTotal":86},{"number":50,"slug":"support-studio","title":"support studio — the reply drafter you can trust","blurb":"how a working support agent runs AI: a policy pack the drafter can't talk around, a banned-promises linter that holds every draft, golden eval sets that measure each macro against fixture cases without authorizing deployment one intent at a time, and the escalation line where money, law, and anger stay human.","status":"live","lessonCount":3,"stepCount":23,"liveLessonCount":3,"liveStepCount":23,"estMinutes":61,"xpTotal":71},{"number":51,"slug":"copy-studio","title":"copy studio — voice from evidence, variants at scale, claims that clear","blurb":"a client's voice measured from their own published copy, variant batches that pass the voice gate or die with a named reason, and a claim check that blocks the ship button while any fact is unsourced. corrections-log culture, as code.","status":"live","lessonCount":3,"stepCount":24,"liveLessonCount":3,"liveStepCount":24,"estMinutes":66,"xpTotal":72},{"number":52,"slug":"analytics-studio","title":"analytics studio — messy export to defensible memo","blurb":"how a working analyst runs AI: a cleaning log that accounts for every dropped row, sample-first checks that catch the plausible-but-wrong aggregation, a memo where every number carries the address of the cell that made it, SR 11-7-style challenge against the AI's recommendation, and the rerun test — same data, same code, same memo, or it isn't done.","status":"live","lessonCount":3,"stepCount":22,"liveLessonCount":3,"liveStepCount":22,"estMinutes":61,"xpTotal":74},{"number":53,"slug":"delivery-studio","title":"delivery studio — direct the build without becoming the bottleneck","blurb":"how a working delivery lead runs AI builders: specs with acceptance checks written before the build, a review that runs the checks instead of admiring the demo, status roll-ups computed from acceptance records instead of standup vibes, the not-ready call made on evidence you can read out loud to a steering committee, and an hours-and-dollars receipt — before/after on a named function, finance's loaded rate, a named owner, the accepted thing that shipped — small enough to defend in a staff meeting.","status":"live","lessonCount":4,"stepCount":33,"liveLessonCount":4,"liveStepCount":33,"estMinutes":76,"xpTotal":98},{"number":54,"slug":"people-studio","title":"people studio — draft fast, decide human, prove everything","blurb":"how a working hr specialist runs AI: a data firewall that keeps employee records out of prompts by code instead of willpower, recruiting briefs where every claim cites a resume line, policy drafts that cannot publish without a named legal reviewer, and the adverse-action paper trail that survives a discrimination claim.","status":"live","lessonCount":3,"stepCount":25,"liveLessonCount":3,"liveStepCount":25,"estMinutes":58,"xpTotal":74},{"number":55,"slug":"ops-studio","title":"ops studio — tribal knowledge to running automation","blurb":"how a working ops lead gets the process out of one person's head and into automation that can be trusted: SOPs captured as versioned data with the invented steps hunted down, recurring checklists with an explicit auto-run vs named-approval line, exception queues instead of silent failures, a receipt against the export for every number in the weekly report, and a PII proxy that masks member identifiers before the model and restores them only for the named human who sends.","status":"live","lessonCount":4,"stepCount":33,"liveLessonCount":4,"liveStepCount":33,"estMinutes":82,"xpTotal":98},{"number":56,"slug":"legal-studio","title":"legal studio — the cite-check gate and the log that survives","blurb":"how a working lawyer runs AI: a citation-verification drill where every authority gets opened at the source (the fabricated ones always look right), clause extraction that's worthless without a pinpoint, privilege rules written as a routing table instead of remembered under deadline, and the review log you'd want in the record when opposing counsel moves for sanctions.","status":"live","lessonCount":3,"stepCount":22,"liveLessonCount":3,"liveStepCount":22,"estMinutes":60,"xpTotal":69},{"number":57,"slug":"agent-studio","title":"agent studio — ship one agent feature before you stand up","blurb":"one sitting, one agent feature, end to end: a tool schema that makes the dangerous call a validation error, a loop that meters its own spend, golden cases wired before the router works, a trace you debug from instead of vibes, and the ratchet that turns tonight's incident into tomorrow's eval. practice at production tempo.","status":"live","lessonCount":3,"stepCount":23,"liveLessonCount":3,"liveStepCount":23,"estMinutes":66,"xpTotal":73},{"number":58,"slug":"fine-tune-a-small-specialist","title":"fine-tune a small specialist","blurb":"one repeated job, a reviewed example set, a QLoRA adapter, and an eval receipt that beats prompt-only on consistency — without claiming the base model got generally smarter.","status":"live","lessonCount":8,"stepCount":28,"liveLessonCount":8,"liveStepCount":28,"estMinutes":130,"xpTotal":65}]},"firstLessonByChapter":{"before-you-build":"from-worker-to-builder","read-the-script":"names-and-values","read-the-data-shapes":"the-shapes-ai-dumps-on-you","read-the-failure":"read-bottom-up","read-the-pipeline":"files-without-fear","mutation-and-state":"why-it-breaks","llm-apis":"before-the-api-what-the-model-is-doing","structured-output":"getting-json-back","mcp":"what-mcp-is","agent-loops":"the-loop","git-and-github":"the-three-states","secrets-and-env":"keeping-keys-safe","prompting":"builder-briefs-before-prompts","agent-traces":"reading-the-trace","evals":"check-before-you-trust","context-and-retrieval":"ask-the-archive","production-tradeoffs":"prompt-caching-correctly","debugging-output":"read-the-trace-not-the-chat","capstone":"pick-a-project-that-ships","agent-harnesses":"project-workspaces-not-product-names","ai-image-generation":"the-image-model-landscape","ai-video-generation":"the-video-model-lineup","programmatic-design":"why-programmatic","harness-engineering":"the-harness-engineering-mindset","intro-to-terminal":"what-the-terminal-is","intro-to-claude-cli":"what-the-claude-cli-is","intro-to-codex-cli":"another-tool-same-idea","team-skills":"what-a-skill-is","dataframes-numpy-pandas":"ml-package-map","sql-for-ml-datasets":"select-filter-aggregate","dataset-pipelines":"csv-jsonl-parquet-tradeoffs","ml-math-and-stats":"vectors-matrices-dot-products","supervised-learning-workflows":"labels-features-splits","unsupervised-learning-and-embeddings":"clustering-without-labels","metrics-and-error-analysis":"confusion-matrix","pytorch-tensors-and-autograd":"tensors-and-shapes","training-loops-and-optimizers":"loss-forward-backward-step","deep-learning-architectures":"cnns-and-local-patterns","feature-experiments-registries":"feature-pipeline-contracts","model-serving-and-mlops":"fastapi-inference-shape","monitoring-cloud-portfolio":"structured-logs-and-alerts","graphic-design-studio":"prompting-like-an-art-director","campaign-studio":"voice-pack-as-data","support-studio":"the-policy-pack","copy-studio":"the-voice-fingerprint","analytics-studio":"the-cleaning-log","delivery-studio":"the-spec-is-the-build","people-studio":"the-data-firewall","ops-studio":"the-sop-is-data","legal-studio":"the-cite-check-gate","agent-studio":"the-tool-contract","fine-tune-a-small-specialist":"the-specialist-not-the-genius"},"detail":{"number":49,"slug":"campaign-studio","title":"campaign studio: one campaign, end to end, with the gate on","blurb":"Run the Fieldnote fall relaunch end to end: copy checked against the voice pack, a batch plan, a calendar where every post has a named approver, a results report where every number traces to the export, and a clip plan that turns one webinar into short clips you can post. Built for the person who owns the send button.","overview":"$2c","status":"live","lessons":[{"slug":"voice-pack-as-data","title":"The voice pack is data — on-voice copy at batch volume","estMinutes":21,"prerequisites":[],"status":"live","steps":[{"type":"read","id":"campaign-studio/voice-pack-as-data/s01-01-the-campaign-and-the-gate","xp":1,"hint":[],"commonMistakes":[],"personalize":false,"phase":"warmup","estSeconds":200,"concept":"campaign-studio-setup","body":"$2d","cta":"Got it","runnable":true},{"type":"read","id":"campaign-studio/voice-pack-as-data/s02-02-lint-the-draft","xp":1,"hint":[],"commonMistakes":[],"personalize":false,"phase":"warmup","estSeconds":170,"concept":"voice-linting","body":"$2e","cta":"Got it","code":"PACK = {\n \"banned\": [\"best-in-class\", \"revolutionary\", \"game-changing\", \"#1\"],\n \"max_words\": 20,\n}\n\ndraft = \"Our revolutionary subscription is the best-in-class way to buy coffee.\"\n\nproblems = []\ntext = draft.lower()\nfor phrase in PACK[\"banned\"]:\n if phrase in text:\n problems.append(f\"banned: {phrase}\")\nif len(draft.split()) > PACK[\"max_words\"]:\n problems.append(\"too long\")\n\nprint(f\"{len(problems)} violation(s)\")\nfor p in problems:\n print(f\"- {p}\")\n","runnable":true},{"type":"predict","id":"campaign-studio/voice-pack-as-data/s03-03-predict-the-violations","xp":3,"hint":[{"level":1,"body":"Check each of the four banned phrases against the draft. best-in-class is absent. Pack order puts revolutionary before game-changing before #1.","cost":0},{"level":2,"body":"Three hits: violations: 3, then - revolutionary, - game-changing, - #1.","cost":0}],"commonMistakes":[],"personalize":false,"phase":"build","estSeconds":70,"concept":"banned-phrase-scan","optional":true,"code":"# Optional observability — a pack scan a teammate might automate.\n# Purpose: walk banned phrases against the draft and list hits in pack order.\n# Gloss: `b in draft.lower()` means \"banned phrase appears in the draft (ignore case).\"\n# Role: read-only example. You pass by predicting the job output, not by writing this.\nBANNED = [\"best-in-class\", \"revolutionary\", \"game-changing\", \"#1\"]\n\ndraft = \"Our revolutionary roaster is the #1 choice for game-changing mornings.\"\n\nhits = [b for b in BANNED if b in draft.lower()]\nprint(f\"violations: {len(hits)}\")\nfor b in hits:\n print(f\"- {b}\")\n","prompt":"Unlock skill: spot which banned phrases an on-voice gate would flag\nbefore Dana sees the subject line.\n\nFieldnote's fall subject draft contains \"revolutionary,\" \"game-changing,\"\nand \"#1\". The pack bans those (plus best-in-class, which is absent).\nHits print in **pack list order**, not sentence order. What does the\nvoice gate report?\n","grader":{"kind":"stdout-equality","expected":"violations: 3\n- revolutionary\n- game-changing\n- #1\n","normalize":"collapse-trailing-newline"}},{"type":"mc","id":"campaign-studio/voice-pack-as-data/s04-04-testable-or-mood","xp":2,"hint":[{"level":1,"body":"The test for a testable rule: could a checklist fail a draft against it without arguing about taste? Exactly one candidate survives — it involves counting words.","cost":0}],"commonMistakes":[],"personalize":false,"phase":"build","estSeconds":60,"concept":"testable-vs-mood-rules","prompt":"You're migrating Fieldnote's brand PDF into the voice pack. Only rules\na script can check go in the `rules` list — the rest either become\nbanned phrases or stay as human-judgment notes. Which candidate\nbelongs in `rules`?\n","options":[{"id":"under_20","label":"2 — sentences under 20 words"},{"id":"warm_prof","label":"1 — warm but professional","explain":"Warmth isn't measurable by a script. This is a mood word — it can live in the prompt to steer the draft, but it can't gate anything, because no check can fail it."},{"id":"punchy","label":"3 — punchy, thumb-stopping copy","explain":"'Punchy' has no test. Ask three marketers whether a line is punchy and you'll get four answers. Mood words go in the prompt, not the rules list."},{"id":"friend","label":"4 — sounds like a knowledgeable friend","explain":"A persona note, and a useful one — for the `reader` field or the prompt. But there's no substring match or word count that verifies 'knowledgeable friend.' Human judgment territory."},{"id":"bold","label":"5 — authentically bold","explain":"Two mood words stapled together. 'Authentic' may be the least testable word in marketing — no substring match, count, or regex measures it. Prompt material at best; it can't gate anything."}],"answerIds":["under_20"],"shuffle":false,"code":"# five lines pulled from Fieldnote's old brand PDF.\ncandidates = [\n \"warm but professional\",\n \"sentences under 20 words\",\n \"punchy, thumb-stopping copy\",\n \"sounds like a knowledgeable friend\",\n \"authentically bold\",\n]\nfor i, c in enumerate(candidates, 1):\n print(f\"{i}. {c}\")\n","runnable":true},{"type":"fill","id":"campaign-studio/voice-pack-as-data/s05-05-fill-the-linter","xp":3,"hint":[{"level":1,"body":"The report Dana reads says banned: revolutionary — not mood: or voice:. The blank is that label word.","cost":0},{"level":2,"body":"`banned` — so each hit becomes `banned: `.","cost":0}],"commonMistakes":[],"personalize":false,"phase":"build","estSeconds":80,"concept":"substring-banned-check","optional":true,"prompt":"Unlock skill: name the **violation label** the voice pack raises on a\ndraft — the string Dana sees on the revision queue.\n\nDraft A: \"A revolutionary blend with true synergy\" → two banned hits.\nDraft B: \"Fresh beans, every Tuesday\" → clean.\n\nThe helper below is complete. Fill the blank with the label prefix the\npack uses for banned hits (the word before the colon in\n`banned: revolutionary`).\n\nExpected output:\n```\n['banned: revolutionary', 'banned: synergy']\n[]\n```\n","code":"# Purpose: attach a labeled violation when a banned phrase appears.\n# Walkthrough: lowercase the draft → for each banned phrase → if present, append \"banned: …\".\n# Gloss: `phrase in text` = substring check. Role: required judgment payload is the blank label.\nPACK = {\n \"banned\": [\"best-in-class\", \"revolutionary\", \"synergy\"],\n \"max_words\": 20,\n}\nLABEL = \"___\" # the prefix before \": phrase\" on each banned hit\n\ndef lint(line):\n problems = []\n text = line.lower()\n for phrase in PACK[\"banned\"]:\n if phrase in text:\n problems.append(f\"{LABEL}: {phrase}\")\n if len(line.split()) > PACK[\"max_words\"]:\n problems.append(\"too long\")\n return problems\n\nprint(lint(\"A revolutionary blend with true synergy\"))\nprint(lint(\"Fresh beans, every Tuesday\"))\n","blanks":[{"id":"banned_label","accept":["banned"],"caseSensitive":false,"normalize":"trim"}]},{"type":"read","id":"campaign-studio/voice-pack-as-data/s06-06-variants-as-records","xp":1,"hint":[],"commonMistakes":[],"personalize":false,"phase":"build","estSeconds":180,"concept":"copy-variants-as-records","body":"$2f","cta":"Got it","code":"# one drafted variant. not a paragraph in a doc — a record.\nvariant = {\n \"id\": \"E3\",\n \"channel\": \"email\",\n \"slot\": \"subject-line\",\n \"text\": \"Your Tuesday beans are on the way\",\n \"cta\": \"none\",\n \"source\": \"ai-draft\",\n \"lint\": [], # filled by the linter\n \"approved_by\": \"\", # filled by a human, never by a script\n}\n\nfor field in (\"id\", \"channel\", \"slot\", \"text\"):\n print(f\"{field:<9} {variant[field]}\")\nprint(f\"lint {'clean' if not variant['lint'] else variant['lint']}\")\n","runnable":true},{"type":"write","id":"campaign-studio/voice-pack-as-data/s07-07-write-lint-copy","xp":5,"hint":[{"level":1,"body":"V2 contains the banned phrase game-changing. V3 is 15 words against a max of 12 — that is too long. V1 and V4 clear both checks. Set result to FAIL and fill the matching violation labels; leave PASS rows with an empty violations list.","cost":0},{"level":2,"body":"verdicts should be:\nV1 PASS []; V2 FAIL [\"banned: game-changing\"]; V3 FAIL [\"too long\"]; V4 PASS [].\nOptional: call lint_copy on a draft to confirm — your verdicts still drive the printed lines.\n","cost":0}],"commonMistakes":[],"personalize":false,"phase":"build","estSeconds":240,"concept":"write-lint-copy","optional":true,"prompt":"Unlock skill: decide whether each Fieldnote subject draft is **on-voice**\nusing testable pack rules (banned list + max words) — not mood adjectives.\n\nDana set `max_words` to 12 for this subject-line slot. The pack is data:\nthe check reads whatever cap the pack states.\n\nFor each of the four drafts, mark PASS or FAIL. On FAIL, name the\nviolation label(s) in pack order:\n- banned phrases first, as `banned: `\n- then `too long` if word count exceeds `max_words`\n\nEdit only the `verdicts` dict in the starter. The helper `lint_copy` is\nalready complete — optional tool if you want a second opinion. Your\nmarks are what print and what the grader checks.\n\nExpected output:\n```\nV1: PASS\nV2: FAIL (banned: game-changing)\nV3: FAIL (too long)\nV4: PASS\n```\n","starter":"$30","grader":{"kind":"stdout-equality","expected":"V1: PASS\nV2: FAIL (banned: game-changing)\nV3: FAIL (too long)\nV4: PASS\n","normalize":"collapse-trailing-newline"},"solution":"$31","hiddenTests":[]},{"type":"checkpoint","id":"campaign-studio/voice-pack-as-data/s08-08-checkpoint","xp":8,"hint":[{"level":1,"body":"Unlock skill: build the **on-voice queue for Dana** — which variants leave the lint gate, which stay in revision, and what to tell the model.","cost":0},{"level":2,"body":"When it runs, stdout should look like: `E1 email PASS …`.","cost":0},{"level":3,"body":"Near-solution lines:\n \"max_words\": 10,\n \"E2\": {\"call\": \"HOLD\", \"violations\": [\"banned: revolutionary\"]},\n \"S2\": {\"call\": \"HOLD\", \"violations\": [\"banned: best-in-class\", \"banned: #1\", \"too long\"]},\n","cost":0}],"commonMistakes":[],"personalize":false,"phase":"check","estSeconds":280,"concept":"on-voice-batch-pass","optional":true,"prompt":"Unlock skill: build the **on-voice queue for Dana** — which variants\nleave the lint gate, which stay in revision, and what to tell the model.\n\nFive Fieldnote relaunch variants (E1/E2/S1/S2/A1) + PACK. For each:\n- SHIP → on-voice; goes to Dana\n- HOLD → off-voice; stay in revision\n\nOn HOLD, list the violation labels the pack would raise (banned phrases\nin list order, then too long). Final line: kept k/5 — batch goes to Dana.\n\nEdit only the `verdicts` dict. Helpers are complete and optional.\nPass = correct ship/hold decisions (3 kept), not writing lint loops.\n\nExpected output:\n```\nE1 email PASS\nE2 email FAIL banned: revolutionary\nS1 social PASS\nS2 social FAIL banned: best-in-class; banned: #1; too long\nA1 ads PASS\nkept 3/5 — batch goes to Dana for approval\n```\n","starter":"$32","grader":{"kind":"stdout-equality","expected":"E1 email PASS\nE2 email FAIL banned: revolutionary\nS1 social PASS\nS2 social FAIL banned: best-in-class; banned: #1; too long\nA1 ads PASS\nkept 3/5 — batch goes to Dana for approval\n","normalize":"collapse-trailing-newline"},"solution":"$33","passThreshold":0.66}],"xpTotal":24},{"slug":"batch-plan-and-calendar","title":"Batch plan and calendar — creative at volume, gated before it ships","estMinutes":19,"prerequisites":["voice-pack-as-data"],"status":"live","steps":[{"type":"read","id":"campaign-studio/batch-plan-and-calendar/s01-01-slots-before-prompts","xp":1,"hint":[],"commonMistakes":[],"personalize":false,"phase":"warmup","estSeconds":190,"concept":"creative-batch-planning","body":"$34","cta":"Got it","runnable":true},{"type":"mc","id":"campaign-studio/batch-plan-and-calendar/s02-02-which-field-is-missing","xp":2,"hint":[{"level":1,"body":"Story frames in the wild usually carry a headline, so an unbriefed model helpfully invents one — in letters from no alphabet. Which field in lesson 1's batch plan made that decision explicit, even when the answer was 'none'?","cost":0}],"commonMistakes":[],"personalize":false,"phase":"build","estSeconds":70,"concept":"generation-brief-fields","prompt":"The story batch comes back and half the variants have garbled\npseudo-words baked into the image — melted letterforms where a\nheadline would go. The model was left to guess about something the\nbrief should have pinned down. Which missing field caused the rework?\n","options":[{"id":"text_on_asset","label":"text_on_asset — the exact words to render, or the explicit value \"none\""},{"id":"subject","label":"subject — the model didn't know what to draw","explain":"The subject is right there in the brief — box, counter, morning light — and the variants show it. The failure is specifically the invented text, and no field in this brief says whether text belongs on the asset at all."},{"id":"size","label":"size — without dimensions the model picked a random crop","explain":"1080x1920 is in the brief, and the variants came back vertical. Garbled letterforms aren't a crop problem — they're what image models produce when they decide, unbriefed, that a story frame usually has a headline."},{"id":"brand_constraints","label":"brand_constraints — the palette rules were too loose","explain":"The constraints line is present and reasonably tight. Palette drift would show as off-brand colors, not melted fake words. The text decision was never made, so the model made it."}],"answerIds":["text_on_asset"],"shuffle":false,"code":"# the brief that went to the image model for the story slot.\nbrief = {\n \"slot\": \"story\",\n \"size\": \"1080x1920\",\n \"subject\": \"the fall subscription box on a kitchen counter, morning light\",\n \"composition\": \"vertical, product in lower half, negative space top third\",\n \"brand_constraints\": \"warm palette, no faces, no competitor mugs\",\n}\nfor k, v in brief.items():\n print(f\"{k}: {v}\")\n","runnable":true},{"type":"write","id":"campaign-studio/batch-plan-and-calendar/s03-03-write-batch-plan","xp":4,"hint":[{"level":1,"body":"Counts come from the read: 6 / 8 / 8 / 10. Line cost is variants times 0.04. Totals: 32 assets and $1.28. When budget halves, cut email-header first — it is decorative; keep ad-banner depth because it carries the media budget.","cost":0},{"level":2,"body":"Fill plan rows to match the expected four lines, set total_assets=32 and total_cost=1.28, cut_first=\"email-header\", cut_why=\"decorative; ad-banner carries the budget\".\n","cost":0}],"commonMistakes":[],"personalize":false,"phase":"build","estSeconds":220,"concept":"batch-plan-cost","optional":true,"prompt":"Unlock skill: turn a vague \"make some assets\" brief into a **concrete\nslot plan + review-time/cost budget** before anyone opens an image tool.\n\nFieldnote relaunch slots (already decided in the read):\nemail-header x6, story x8, feed-square x8, ad-banner x10.\nCost per image: $0.04. Generation dollars are tiny; the number that\nscales is 32 variants a human must review before the calendar gate.\n\nFill the `plan` table (slot, variants, line_cost). Confirm totals.\nThen answer the tradeoff: if budget halves, which slot do you cut first\nand why? Put that in `cut_first` / `cut_why` (printed as a note line —\ngrader keys on the plan totals and slot lines).\n\nEdit only the JUDGMENT block. The helper `batch_plan` is optional.\n\nExpected output:\n```\nemail-header x6 $0.24\nstory x8 $0.32\nfeed-square x8 $0.32\nad-banner x10 $0.40\ntotal: 32 assets, $1.28\ncut-first: email-header — decorative; ad-banner carries the budget\n```\n","starter":"$35","grader":{"kind":"stdout-equality","expected":"email-header x6 $0.24\nstory x8 $0.32\nfeed-square x8 $0.32\nad-banner x10 $0.40\ntotal: 32 assets, $1.28\ncut-first: email-header — decorative; ad-banner carries the budget\n","normalize":"collapse-trailing-newline"},"solution":"$36","hiddenTests":[]},{"type":"read","id":"campaign-studio/batch-plan-and-calendar/s04-04-the-calendar-is-data","xp":1,"hint":[],"commonMistakes":[],"personalize":false,"phase":"build","estSeconds":190,"concept":"calendar-approval-column","body":"$37","cta":"Got it","code":"# two-week relaunch calendar, first rows. timezone: America/Chicago.\nrows = [\n {\"date\": \"09-02\", \"channel\": \"email\", \"piece\": \"launch-email\",\n \"owner\": \"you\", \"approved_by\": \"dana\", \"status\": \"approved\"},\n {\"date\": \"09-02\", \"channel\": \"social\", \"piece\": \"teaser-story\",\n \"owner\": \"you\", \"approved_by\": \"\", \"status\": \"approved\"},\n {\"date\": \"09-03\", \"channel\": \"social\", \"piece\": \"meme-post\",\n \"owner\": \"you\", \"approved_by\": \"dana\", \"status\": \"rejected\"},\n]\n\nfor r in rows:\n gate = \"OK\" if r[\"status\"] == \"approved\" and r[\"approved_by\"] else \"HOLD\"\n print(f\"{r['date']} {r['channel']:<6} {r['piece']:<13} {gate}\")\n","runnable":true},{"type":"predict","id":"campaign-studio/batch-plan-and-calendar/s05-05-predict-the-queue","xp":3,"hint":[{"level":1,"body":"launch-email: approved + dana → schedule. teaser-story: approved but no name → out. meme-post: rejected → out. ad-set-a: approved + sam → schedule.","cost":0},{"level":2,"body":"scheduled: 2, then 09-02 launch-email and 09-04 ad-set-a.","cost":0}],"commonMistakes":[],"personalize":false,"phase":"build","estSeconds":80,"concept":"approval-gate-filter","optional":true,"code":"# Optional observability — calendar gate as a teammate might code it.\n# Purpose: keep only rows with status approved AND a named approver.\n# Gloss: empty approved_by (\"\") means nobody put their name on the row.\n# Role: read-only. Pass = predicting SCHEDULE survivors in ops language.\nrows = [\n {\"date\": \"09-02\", \"piece\": \"launch-email\", \"approved_by\": \"dana\", \"status\": \"approved\"},\n {\"date\": \"09-02\", \"piece\": \"teaser-story\", \"approved_by\": \"\", \"status\": \"approved\"},\n {\"date\": \"09-03\", \"piece\": \"meme-post\", \"approved_by\": \"dana\", \"status\": \"rejected\"},\n {\"date\": \"09-04\", \"piece\": \"ad-set-a\", \"approved_by\": \"sam\", \"status\": \"approved\"},\n]\n\nqueue = [r for r in rows if r[\"status\"] == \"approved\" and r[\"approved_by\"]]\n\nprint(f\"scheduled: {len(queue)}\")\nfor r in queue:\n print(f\"{r['date']} {r['piece']}\")\n","prompt":"Unlock skill: predict which calendar rows **SCHEDULE** — approved with\na named human — vs which stay out (nameless \"approved\" or rejected).\n\nFour relaunch rows. Teaser-story has status approved but empty\napproved_by — that is the trap. What prints for the schedule queue?\n","grader":{"kind":"stdout-equality","expected":"scheduled: 2\n09-02 launch-email\n09-04 ad-set-a\n","normalize":"collapse-trailing-newline"}},{"type":"fill","id":"campaign-studio/batch-plan-and-calendar/s06-06-fill-the-gate","xp":3,"hint":[{"level":1,"body":"Policy must be on. True means teaser-story (empty approved_by) is held; only dana and sam rows reach the queue.","cost":0},{"level":2,"body":"`True` — queue drops to 2 rows: launch-email and winback.","cost":0}],"commonMistakes":[],"personalize":false,"phase":"build","estSeconds":110,"concept":"fix-approval-gate","optional":true,"prompt":"Unlock skill: enforce the policy that a row needs **both** approved\nstatus and a named approver before it reaches the publish queue.\n\nThe teaser-story row is approved with an empty approved_by. A leaky\ngate that checks status alone would schedule it. Set the policy\nconstant so nameless approved rows are held.\n\nFill `REQUIRE_NAMED_APPROVER` with True (policy on) or False (leaky).\n\nExpected output:\n```\nqueue: 2 rows\n09-02 launch-email -> approved by dana\n09-05 winback -> approved by sam\n```\n","code":"$38","blanks":[{"id":"require_named","accept":["True","true","1"],"caseSensitive":false,"normalize":"trim"}]},{"type":"checkpoint","id":"campaign-studio/batch-plan-and-calendar/s07-07-checkpoint","xp":8,"hint":[{"level":1,"body":"Unlock skill: nothing enters the scheduler without **status=approved AND a named approver**; rejected rows stay as the record of \"no\". Calendar triage (HITL) on five relaunch ro…","cost":0},{"level":2,"body":"When it runs, stdout should look like: `SCHEDULE 09-02 email launch-email (dana) …`.","cost":0},{"level":3,"body":"Near-solution lines:\n \"teaser-story\": \"BLOCKED\",\n \"meme-post\": \"KEPT-REJECTED\",\n","cost":0}],"commonMistakes":[],"personalize":false,"phase":"check","estSeconds":280,"concept":"full-calendar-gate","optional":true,"prompt":"Unlock skill: nothing enters the scheduler without **status=approved\nAND a named approver**; rejected rows stay as the record of \"no\".\n\nCalendar triage (HITL) on five relaunch rows. For each piece, set a\nroute in `verdicts`:\n- SCHEDULE — approved + non-empty approved_by\n- BLOCKED — approved but no approver (teaser-story failure mode)\n- KEPT-REJECTED — status rejected; stays in the file\n\nEdit only the JUDGMENT block. Optional `gate` helper is complete.\nPass = correct bucket for every row + summary counts.\n\nExpected output:\n```\nSCHEDULE 09-02 email launch-email (dana)\nSCHEDULE 09-04 ads ad-set-a (sam)\nSCHEDULE 09-05 email winback (sam)\nBLOCKED 09-02 social teaser-story (no approver)\nKEPT 09-03 social meme-post (rejected, stays in the file)\n3 scheduled, 1 blocked, 1 rejected-and-kept\n```\n","starter":"$39","grader":{"kind":"stdout-equality","expected":"SCHEDULE 09-02 email launch-email (dana)\nSCHEDULE 09-04 ads ad-set-a (sam)\nSCHEDULE 09-05 email winback (sam)\nBLOCKED 09-02 social teaser-story (no approver)\nKEPT 09-03 social meme-post (rejected, stays in the file)\n3 scheduled, 1 blocked, 1 rejected-and-kept\n","normalize":"collapse-trailing-newline"},"solution":"$3a","passThreshold":0.66}],"xpTotal":22},{"slug":"the-numbers-pass","title":"The numbers pass: every claimed metric traced to the export","estMinutes":23,"prerequisites":["voice-pack-as-data","batch-plan-and-calendar"],"status":"live","steps":[{"type":"read","id":"campaign-studio/the-numbers-pass/s01-01-a-report-is-a-draft-over-data","xp":1,"hint":[],"commonMistakes":[],"personalize":false,"phase":"warmup","estSeconds":190,"concept":"report-over-export","body":"$3b","cta":"Got it","runnable":true},{"type":"predict","id":"campaign-studio/the-numbers-pass/s02-02-predict-the-verdict","xp":3,"hint":[{"level":1,"body":"618 / 18200 * 100 ≈ 3.4% at one decimal. Claimed 3.6% does not match → invented.","cost":0},{"level":2,"body":"claimed CTR: 3.6%, export CTR: 3.4% (two spaces after the colon), verdict: invented.","cost":0}],"commonMistakes":[],"personalize":false,"phase":"warmup","estSeconds":70,"concept":"trace-one-claim","optional":true,"code":"# Optional recompute tool — CTR from export cells.\n# Purpose: recompute claimed CTR from clicks/sent and mark traced vs invented.\n# Gloss: actual:.1f formats one decimal place. Role: observability; score the verdict call.\nexport = {\"sent\": 18200, \"opens\": 4914, \"clicks\": 618}\n\nclaimed = \"3.6%\"\nactual = export[\"clicks\"] / export[\"sent\"] * 100\n\nprint(f\"claimed CTR: {claimed}\")\nprint(f\"export CTR: {actual:.1f}%\")\nprint(\"verdict:\", \"traced\" if f\"{actual:.1f}%\" == claimed else \"invented\")\n","prompt":"Unlock skill: decide whether a claimed CTR is **traced** to the export\nor **invented** under social pressure.\n\nThe AI summary claims 3.6% CTR. Recompute from export (clicks / sent —\nthis platform's convention) at one decimal. 618 / 18200 ≈ 3.4%. What\nprints — including the verdict call?\n","grader":{"kind":"stdout-equality","expected":"claimed CTR: 3.6%\nexport CTR: 3.4%\nverdict: invented\n","normalize":"collapse-trailing-newline"}},{"type":"write","id":"campaign-studio/the-numbers-pass/s03-03-write-verify-numbers","xp":5,"hint":[{"level":1,"body":"Compare each claim value to the export cell for that metric. Clicks claim 655 but export clicks is 618 — INVENTED. The other three match. One invented claim means the report is NOT approval-ready.","cost":0},{"level":2,"body":"Set clicks mark to INVENTED with export_actual 618; leave the other three TRACED; set approval_ready = False.\n","cost":0}],"commonMistakes":[],"personalize":false,"phase":"build","estSeconds":240,"concept":"write-verify-numbers","optional":true,"prompt":"Unlock skill: every figure in an AI-drafted report is **TRACED to the\nexport** or it is fiction — catch invented metrics before approval.\n\nSide-by-side: claims vs export. For each claim mark TRACED or INVENTED.\nOn INVENTED, record the export actual. Then decide: report\napproval-ready? Y/N (only if every claim traced).\n\nEdit only the JUDGMENT block. Optional `verify_numbers` helper is\ncomplete. No f-string formatter authorship required.\n\nExpected output:\n```\nemails sent 18200 TRACED\nopens 4914 TRACED\nclicks 655 INVENTED (export: 618)\nunsubscribes 43 TRACED\n3/4 traced — report is NOT approval-ready\n```\n","starter":"$3c","grader":{"kind":"stdout-equality","expected":"emails sent 18200 TRACED\nopens 4914 TRACED\nclicks 655 INVENTED (export: 618)\nunsubscribes 43 TRACED\n3/4 traced — report is NOT approval-ready\n","normalize":"collapse-trailing-newline"},"solution":"$3d","hiddenTests":[]},{"type":"read","id":"campaign-studio/the-numbers-pass/s04-04-denominators-and-the-ftc","xp":1,"hint":[],"commonMistakes":[],"personalize":false,"phase":"build","estSeconds":240,"concept":"denominators-and-banned-outright","body":"$3e","cta":"Got it","code":"# two real 2026 statistics about AI in marketing. both true. not comparable.\nstats = [\n {\"figure\": \"75%\", \"of_what\": \"marketers who have adopted AI at all\",\n \"source\": \"Salesforce State of Marketing 2026 (vendor survey)\"},\n {\"figure\": \"24.2%\", \"of_what\": \"marketing activities AI is used in\",\n \"source\": \"The CMO Survey, Spring 2026 wave\"},\n]\nfor s in stats:\n print(f\"{s['figure']:>6} of {s['of_what']}\")\n print(f\" ({s['source']})\")\n","runnable":true},{"type":"mc","id":"campaign-studio/the-numbers-pass/s05-05-what-may-not-ship","xp":2,"hint":[{"level":1,"body":"Three of these are claims a human can check and sign. One is a testimonial from a person who does not exist — and 16 CFR Part 465 has a specific phrase, and a specific per-violation penalty, for exactly that.","cost":0}],"commonMistakes":[],"personalize":false,"phase":"build","estSeconds":70,"concept":"banned-outright-vs-reviewable","prompt":"Four AI drafts are queued for review. Three can move through the\nstudio's gates — lint, then a named approver. One may not ship at any\napproval level, with sign-off from anyone, ever. Which?\n","options":[{"id":"fake_review","label":"C — the five-star review from 'Sarah K.'"},{"id":"subject","label":"A — the email subject line","explain":"Squarely in AI-may-draft territory: lint it against the voice pack, put it in the calendar, get an approver on the row. That's what the whole pipeline is for."},{"id":"ad_copy","label":"B — 'Roasted weekly in small batches'","explain":"A factual claim, so it needs a fact-check — is the roasting actually weekly? — and an approver. But it's a checkable claim about your own product, which makes it reviewable, not banned."},{"id":"pricing","label":"D — 'The most affordable subscription in the US'","explain":"Close — an unsubstantiated superlative is deceptive if false, and this one probably dies in review. But it dies BY review: it needs substantiation and a named approver, and with real price data behind it, a version of it could ship. A category apart from the fabricated reviewer."}],"answerIds":["fake_review"],"shuffle":false,"code":"# four pieces the AI drafted for relaunch week.\ndrafted = [\n \"A: email subject — 'Your Tuesday beans, now with a fall tier'\",\n \"B: ad copy — 'Roasted weekly in small batches'\",\n \"C: five-star review from 'Sarah K., Austin' — 'Fieldnote changed my mornings!'\",\n \"D: pricing claim — 'The most affordable subscription in the US'\",\n]\nfor line in drafted:\n print(line)\n","runnable":true},{"type":"fill","id":"campaign-studio/the-numbers-pass/s06-06-fill-the-policy-router","xp":3,"hint":[{"level":1,"body":"Unlisted pieces (case study) must not silently become ai_may_draft. The safe default is a named human signs.","cost":0},{"level":2,"body":"`named_approver` — ad copy hits the first tier, testimonial hits banned_outright, case study falls to the default.","cost":0}],"commonMistakes":[],"personalize":false,"phase":"build","estSeconds":80,"concept":"policy-router-membership","optional":true,"prompt":"Unlock skill: route each piece type to its **policy tier** —\nai_may_draft, named_approver, or banned_outright — not expression golf.\n\nThe helper walks tiers in order and returns the first match; unlisted\npieces default to named_approver (when in doubt, a human signs).\n\nFill the blank with the safe default tier for pieces that appear on\nno list (e.g. \"case study\").\n\nExpected output:\n```\nai_may_draft\nbanned_outright\nnamed_approver\n```\n","code":"# Purpose: send each drafted piece type to its brand-safety tier.\n# Walkthrough: check ai_may_draft list → named_approver → banned_outright → default.\n# Gloss: `piece in kinds` means this piece type is on that tier's list.\n# Role: judgment payload = the default tier name when unlisted.\nPOLICY = {\n \"ai_may_draft\": [\"social post\", \"email subject\", \"ad copy\"],\n \"named_approver\": [\"press quote\", \"pricing claim\", \"partner mention\"],\n # Fictional brand policy: do not invent customer experiences.\n # This fixture bans these generated artifact types; it is not a legal test\n # of whether editing a genuine customer account with AI makes it fabricated.\n \"banned_outright\": [\"customer review\", \"testimonial\"],\n}\n\nDEFAULT_TIER = \"___\" # when a piece is on no list\n\ndef route(piece):\n for tier, kinds in POLICY.items():\n if piece in kinds:\n return tier\n return DEFAULT_TIER\n\nprint(route(\"ad copy\"))\nprint(route(\"testimonial\"))\nprint(route(\"case study\"))\n","blanks":[{"id":"default_tier","accept":["named_approver"],"caseSensitive":false,"normalize":"trim"}]},{"type":"checkpoint","id":"campaign-studio/the-numbers-pass/s07-07-checkpoint","xp":8,"hint":[{"level":1,"body":"Do **not** write `def campaign_audit`.","cost":0},{"level":2,"body":"When it runs, stdout should look like: `sent 18200 TRACED …`.","cost":0},{"level":3,"body":"Near-solution lines:\n \"signups\": {\"mark\": \"INVENTED\", \"export_actual\": 212},\n \"we-hit-record post\": {\"lane\": \"BLOCKED\", \"name\": \"\"},\npublish_verdict = \"HELD (1 invented number(s), 1 unapproved row(s))\"\n","cost":0}],"commonMistakes":[],"personalize":false,"phase":"build","estSeconds":290,"concept":"full-campaign-audit","optional":true,"prompt":"Unlock skill: **publish judgment under dual gate** — clear only if\n(1) every claim traces to export AND (2) every queue piece has a named\napprover; otherwise HOLD with invented-count and blocked-count.\n\nPacket (fixture, not code homework):\n- export cells vs report claims (signups 230 vs export 212 is planted)\n- queue: results-recap email OK/dana; we-hit-record post empty approver\n\n1. Numbers lane — mark each claim TRACED or INVENTED (export actual if invented)\n2. Approver lane — OK(name) or BLOCKED(no approver) per queue row\n3. Publish call — CLEARED FOR PUBLISH or HELD (N invented, M unapproved)\n\nEdit only the JUDGMENT block. Do **not** write `def campaign_audit`.\nWrong CLEARED fails even if you spotted one issue.\n\nExpected output:\n```\nsent 18200 TRACED\nopens 4914 TRACED\nclicks 618 TRACED\nsignups 230 INVENTED (export: 212)\nresults-recap email: OK (dana)\nwe-hit-record post: BLOCKED (no approver)\nverdict: HELD (1 invented number(s), 1 unapproved row(s))\n```\n","starter":"$3f","grader":{"kind":"stdout-equality","expected":"sent 18200 TRACED\nopens 4914 TRACED\nclicks 618 TRACED\nsignups 230 INVENTED (export: 212)\nresults-recap email: OK (dana)\nwe-hit-record post: BLOCKED (no approver)\nverdict: HELD (1 invented number(s), 1 unapproved row(s))\n","normalize":"collapse-trailing-newline"},"solution":"$40","passThreshold":0.66},{"type":"read","id":"campaign-studio/the-numbers-pass/s08-08-take-it-to-your-analyst","xp":1,"hint":[],"commonMistakes":[],"personalize":false,"phase":"build","estSeconds":160,"concept":"hand-the-why-to-the-data-owner","body":"$41","cta":"Got it","runnable":true},{"type":"mc","id":"campaign-studio/the-numbers-pass/s09-09-the-send-time-claim","xp":2,"hint":[{"level":1,"body":"Compare the two rows on everything except the time. If the sends differ in more than one way, the export alone can't say which difference caused the gap.","cost":0},{"level":2,"body":"A good note: Question: should next week's send go out at 8am? File: relaunch export, the two Tuesday sends. Claim: 412 vs 206 clicks on 9,100 each. What would change my mind: the gap shrinks when each list is split across both times.","cost":0}],"commonMistakes":[],"personalize":false,"phase":"check","estSeconds":100,"concept":"pattern-needs-a-data-owner","prompt":"You asked the model what else the relaunch export shows. It came back with: \"Morning sends get twice the clicks. Move every send to 8am.\"\n\nThe relaunch went out in two sends:\n\n| Send | Went to | Sent | Clicks |\n|---|---|---|---|\n| 8am Tuesday | existing subscribers | 9,100 | 412 |\n| 3pm Tuesday | new sign-ups from the fall tier ad | 9,100 | 206 |\n\nThe math checks out: 412 plus 206 is the 618 clicks in the export. Dana wants next week's send time decided by Thursday. What do you do?\n","options":[{"id":"switch","label":"Move next week's send to 8am, since twice the clicks is too big to ignore","explain":"The gap is real, and the two sends went to different lists. People who already subscribe click more at any hour, so this export can't tell you whether the time or the list made the difference."},{"id":"recheck","label":"Ask the AI to double-check its math before Thursday","explain":"The math is already right, and you checked it yourself. The problem is what's being compared, and rechecking the arithmetic won't surface that."},{"id":"note","label":"Send your analyst the four-part note and point out the two different lists"},{"id":"drop","label":"Drop the idea, since the lists were different and the result means nothing","explain":"The result is a weak test, and send time can still matter. Your analyst can set up a fair one by splitting each list across both times next week, in time for Dana's call."}],"answerIds":["note"],"shuffle":false,"runnable":true}],"xpTotal":26},{"slug":"long-form-to-short-clips","title":"One webinar, five clips: long recordings into short ones","estMinutes":22,"prerequisites":["the-numbers-pass"],"status":"live","steps":[{"type":"read","id":"campaign-studio/long-form-to-short-clips/s01-01-one-recording-two-weeks","xp":1,"hint":[],"commonMistakes":[],"personalize":false,"phase":"warmup","estSeconds":180,"concept":"clip-plan-intro","body":"$42","cta":"Got it","runnable":true},{"type":"read","id":"campaign-studio/long-form-to-short-clips/s02-02-the-prompt-and-eight-moments","xp":1,"hint":[],"commonMistakes":[],"personalize":false,"phase":"warmup","estSeconds":240,"concept":"clip-candidates-from-transcript","body":"$43","cta":"Got it","runnable":true},{"type":"mc","id":"campaign-studio/long-form-to-short-clips/s03-03-which-clip-opens","xp":2,"hint":[{"level":1,"body":"Check each candidate against the three limits you were given: under 60 seconds, makes sense on its own, and no number that disagrees with the price sheet.","cost":0}],"commonMistakes":[],"personalize":false,"phase":"build","estSeconds":80,"concept":"pick-the-lead-clip","prompt":"Day 1 of the push gets one clip, and for most viewers it's the first time they'll see Marta. Using the table and Dana's limits, which candidate should open the push?\n","options":[{"id":"c2","label":"C2, the price line: it has the top score, and price is the main reason to try the tier","explain":"The score is high, but Marta says about $6 a bag and the price sheet says $3.50. A clip with a wrong price can't open the push, and it can't run later either."},{"id":"c6","label":"C6, the burned batch story: it ties for the top score and it's the best story Marta told","explain":"It's a strong story, and it runs 90 seconds against Dana's 60 second limit. It can still run later in the push once you trim it, which you'll do in the plan."},{"id":"c5","label":"C5, \"Coffee is a fruit\": it makes sense alone, runs 36 seconds and has no claim to check"},{"id":"c3","label":"C3, the sweeter second cup: it's about taste, which is what the fall tier is selling","explain":"It opens with \"So like I said before\", so a new viewer starts in the middle of a thought. A clip has to make sense to someone who never saw the webinar."}],"answerIds":["c5"],"shuffle":false,"runnable":true},{"type":"read","id":"campaign-studio/long-form-to-short-clips/s04-04-four-checks","xp":1,"hint":[],"commonMistakes":[],"personalize":false,"phase":"build","estSeconds":240,"concept":"four-clip-checks","body":"$44","cta":"Got it","runnable":true},{"type":"mc","id":"campaign-studio/long-form-to-short-clips/s05-05-the-price-line","xp":2,"hint":[{"level":1,"body":"The fix has to work for someone watching with the sound on. Which choice leaves no wrong number anywhere in what Fieldnote posts?","cost":0}],"commonMistakes":[],"personalize":false,"phase":"build","estSeconds":80,"concept":"captions-match-the-audio","prompt":"You like C2. Marta explains the fall tier clearly, and price is the main reason people switch. She says it saves \"about six dollars a bag\", and the price sheet says $16 against $19.50, which is $3.50. What do you do with C2?\n","options":[{"id":"caption","label":"Post C2 with the caption changed to $3.50, so the text on screen shows the right number","explain":"The audio still says six dollars. A caption that disagrees with the speaker looks like a cover-up, and anyone watching with sound on hears the wrong number."},{"id":"hold","label":"Hold C2, give its slot to another clip, and send Marta the right number for next time"},{"id":"footnote","label":"Post C2 with \"prices vary, see site for details\" added as text at the end of the clip","explain":"A footnote doesn't make six dollars true. The numbers pass rule applies to clips too: a number that doesn't trace to a source doesn't ship."},{"id":"voice","label":"Use an AI voice tool to re-record her line saying $3.50, then post C2 on day 1","explain":"That puts words in a real person's mouth that she never said. Even with the right number, you'd need Marta's clear yes first, and you have seven other candidates."}],"answerIds":["hold"],"shuffle":false,"runnable":true},{"type":"mc","id":"campaign-studio/long-form-to-short-clips/s06-06-the-screen-line","xp":2,"hint":[{"level":1,"body":"Three checks: no banned phrase, 10 words or fewer, and every promise in the line is kept by the clip.","cost":0}],"commonMistakes":[],"personalize":false,"phase":"build","estSeconds":80,"concept":"screen-line-in-voice","prompt":"The AI suggested four lines of on-screen text for C1, the clip where Marta says most people ruin good coffee in the last ten seconds of brewing. Her tip lands about 20 seconds into the 39 second clip. Which line can go on screen?\n","options":[{"id":"number_one","label":"\"The #1 mistake that ruins your morning coffee\"","explain":"\"#1\" is on the voice pack's banned list, and nobody measured a ranking of coffee mistakes."},{"id":"baristas","label":"\"What baristas don't want you to know about coffee\"","explain":"Nobody in the clip mentions baristas or a secret. The line promises something the clip doesn't contain, which breaks the rule from the last read."},{"id":"watch_end","label":"\"Stop ruining your coffee and watch until the end\"","explain":"The tip lands about 20 seconds in, so there's nothing waiting at the end. The line asks for attention and gives no information back."},{"id":"plain","label":"\"The last ten seconds of brewing matter most\""}],"answerIds":["plain"],"shuffle":false,"runnable":true},{"type":"read","id":"campaign-studio/long-form-to-short-clips/s07-07-the-posting-plan","xp":1,"hint":[],"commonMistakes":[],"personalize":false,"phase":"build","estSeconds":180,"concept":"clip-posting-plan","body":"$45","cta":"Got it","runnable":true},{"type":"mc","id":"campaign-studio/long-form-to-short-clips/s08-08-ready-for-the-scheduler","xp":3,"hint":[{"level":1,"body":"Look at where each cut starts, then check the AI flag column for that clip in the table from step 2.","cost":0},{"level":2,"body":"C4 still starts at 22:10, which is the moment Marta reads out a viewer's name.","cost":0}],"commonMistakes":[],"personalize":false,"phase":"build","estSeconds":80,"concept":"plan-row-ship-or-hold","prompt":"Here's the draft plan. Captions were checked on all five clips, and each one links to the fall tier page.\n\n- Day 1: C5, 28:44 to 29:20, \"This cup tastes like the fruit it came from\", approved by Dana\n- Day 3: C1, 06:12 to 06:51, \"The last ten seconds of brewing matter most\", approved by Sam\n- Day 6: C8, 44:30 to 45:05, \"Fresh roast every week. Skip a week anytime\", approved by Dana\n- Day 9: C4, 22:10 to 23:02, \"Decaf that still tastes like coffee\", approved by Dana\n- Day 12: C6, 34:22 to 35:07, \"The batch Marta burned, and what it taught her\", approved by Sam\n\nOne row can't go to the scheduler yet. Which one?\n","options":[{"id":"day9","label":"Day 9, the decaf clip"},{"id":"day12","label":"Day 12, the burned batch clip","explain":"C6 was trimmed from 90 seconds to 45 (34:22 to 35:07), so it fits Dana's limit, and Sam approved it."},{"id":"day6","label":"Day 6, the shipping clip","explain":"Monday shipping and skipping a week are both in the brief, so Marta's line is true, and the screen line passes the voice pack."},{"id":"day1","label":"Day 1, the coffee is a fruit clip","explain":"It runs 36 seconds, makes sense on its own, makes no claim to check, and Dana approved it."}],"answerIds":["day9"],"shuffle":false,"runnable":true},{"type":"read","id":"campaign-studio/long-form-to-short-clips/s09-09-your-clip-plan","xp":1,"hint":[],"commonMistakes":[],"personalize":false,"phase":"check","estSeconds":160,"concept":"clip-plan-deliverable","body":"$46","cta":"Got it","runnable":true}],"xpTotal":14}],"liveLessonCount":4,"liveStepCount":33,"xpTotal":86}},"chapter":"$28:props:tree:detail","lesson":"$28:props:tree:detail:lessons:3","step":"$28:props:tree:detail:lessons:3:steps:4","stepIndex":4,"nav":{"tocNext":{"chapterSlug":"campaign-studio","lessonSlug":"long-form-to-short-clips","stepIndex":5},"tocPrev":{"chapterSlug":"campaign-studio","lessonSlug":"long-form-to-short-clips","stepIndex":3},"nextByPath":{"developer":{"chapterSlug":"campaign-studio","lessonSlug":"long-form-to-short-clips","stepIndex":5},"marketer":{"chapterSlug":"campaign-studio","lessonSlug":"long-form-to-short-clips","stepIndex":5},"designer":{"chapterSlug":"campaign-studio","lessonSlug":"long-form-to-short-clips","stepIndex":5},"customer-service":{"chapterSlug":"campaign-studio","lessonSlug":"long-form-to-short-clips","stepIndex":5},"copywriter":{"chapterSlug":"campaign-studio","lessonSlug":"long-form-to-short-clips","stepIndex":5},"data-analyst":{"chapterSlug":"campaign-studio","lessonSlug":"long-form-to-short-clips","stepIndex":5},"project-manager":{"chapterSlug":"campaign-studio","lessonSlug":"long-form-to-short-clips","stepIndex":5},"hr-specialist":{"chapterSlug":"campaign-studio","lessonSlug":"long-form-to-short-clips","stepIndex":5},"operations":{"chapterSlug":"campaign-studio","lessonSlug":"long-form-to-short-clips","stepIndex":5},"lawyer":{"chapterSlug":"campaign-studio","lessonSlug":"long-form-to-short-clips","stepIndex":5}},"prevByPath":{"developer":{"chapterSlug":"campaign-studio","lessonSlug":"long-form-to-short-clips","stepIndex":3},"marketer":{"chapterSlug":"campaign-studio","lessonSlug":"long-form-to-short-clips","stepIndex":3},"designer":{"chapterSlug":"campaign-studio","lessonSlug":"long-form-to-short-clips","stepIndex":3},"customer-service":{"chapterSlug":"campaign-studio","lessonSlug":"long-form-to-short-clips","stepIndex":3},"copywriter":{"chapterSlug":"campaign-studio","lessonSlug":"long-form-to-short-clips","stepIndex":3},"data-analyst":{"chapterSlug":"campaign-studio","lessonSlug":"long-form-to-short-clips","stepIndex":3},"project-manager":{"chapterSlug":"campaign-studio","lessonSlug":"long-form-to-short-clips","stepIndex":3},"hr-specialist":{"chapterSlug":"campaign-studio","lessonSlug":"long-form-to-short-clips","stepIndex":3},"operations":{"chapterSlug":"campaign-studio","lessonSlug":"long-form-to-short-clips","stepIndex":3},"lawyer":{"chapterSlug":"campaign-studio","lessonSlug":"long-form-to-short-clips","stepIndex":3}}}}]

One webinar, five clips: long recordings into short ones — step 5 of 9