How to write better ChatGPT prompts

Better ChatGPT prompts are not magic phrases. They are clear asks: what you want, what the model needs to know, what it must not invent, and how the answer should look. You practice that loop until it becomes muscle memory. That is the company dojo idea in one sentence: you get better by doing the work, not by collecting tip threads.

This guide is for the common learner. You already open ChatGPT at work. The output is uneven. You want fewer dead ends and less editing from scratch.

When you are ready to practice in a guided path, start free in the browser: Chapter 00 preview. No card. No install. No signup.

Why vague prompts fail (and what to fix first)

A weak prompt usually fails for one of four reasons:

1. No outcome. "Help with the launch" is a topic, not a deliverable.
2. Missing context. The model does not know your audience, product, constraints, or what "good" looks like for you.
3. No boundaries. Without limits, models fill gaps with confident guesses.
4. No shape. If you do not specify format, you get a wall of text you then reshape by hand.

You do not fix this by stuffing the prompt with buzzwords. You fix it by naming the job.

Before: Write something about our Q4 campaign.

After: Write a one-page campaign brief for a B2B SaaS product that helps ops teams cut invoice cycle time. Audience: VP Operations. Tone: direct, no hype. Include: problem, offer, three proof points I will fill in later (mark each as [NEED PROOF]), CTA, and open questions. Do not invent customer names or metrics.

The second version is longer. It is also easier to edit, because the model is working inside a box you defined.

The five-part prompt that covers most work

Use this as a default scaffold. Skip a part only when it truly does not apply.

1. Goal (the deliverable)

Say what you want out: a brief, a reply, a checklist, a rewrite, a plan with owners. Prefer nouns you could hand to a coworker.

• "Draft three subject lines under 50 characters."
• "Turn these meeting notes into action items with owners and due dates."
• "Explain this error like I am a junior engineer onboarding to the repo."

2. Context (what only you know)

Paste the facts the model cannot guess: product, audience, channel, constraints, prior decisions, excerpts of real copy or logs. Short beats encyclopedic. If a fact is unknown, say so.

3. Constraints (what not to do)

Name the landmines: no fake stats, no legal advice, no customer PII, stay under a word count, match a voice sample, cite only from the text you pasted.

Constraints are how you stop polite hallucination. Models are fluent. Fluency is not accuracy.

4. Format (the shape of the answer)

Ask for structure up front: bullets, table, JSON fields, email with subject + body, "revise only the weak paragraphs." Format cuts your editing time more than almost any other trick.

5. Iterate (the second message is the skill)

Treat the first reply as a draft. Your follow-up should be specific:

• "Cut the intro. Keep the three options. Make option 2 more skeptical."
• "You invented a metric. Remove it or mark [NEED SOURCE]."
• "Rewrite in the voice of the sample below. Keep the same facts."

People who "get good at ChatGPT" are usually people who got good at revision instructions.

A worked example: from mush to usable

Job: You need an internal update for stakeholders after a delayed launch.

Weak prompt: Write a status update about the delay.

Stronger prompt:

Write an internal status update (150-220 words) for company leadership.

Facts (use only these):
- Feature X slipped from Oct 3 to Oct 17
- Cause: vendor API latency; fix in progress with vendor ticket #4821
- Customer impact: no outage; onboarding demos paused for two accounts
- Next check-in: Friday standup

Audience: execs who want decisions, not theater.
Tone: calm, accountable, no blame theater.
Structure: what changed, why, impact, what we need from leadership (one ask), next date.
Do not invent metrics, customer names, or root causes beyond the facts above.

You can still dislike the draft. That is fine. You now have something you can correct in one pass instead of rebuilding from a foggy paragraph.

Habits that compound faster than "better prompts"

Keep a small library

Save prompts that worked: brief template, meeting-notes-to-actions, review-this-diff, rewrite-in-voice. Reuse beats reinventing. Role paths in the dojo are built around that idea: the same curriculum, routed through the work you already do. Browse routes on the role paths hub.

Separate drafting from verifying

Use the model to draft. You verify claims, numbers, citations, and anything that could embarrass you in front of a customer or counsel. For writing roles, that verification loop is the job; see the copywriter path if that is your lane.

Give the model a role only when it helps

"You are a senior marketer" is weak alone. "You are a senior marketer reviewing a junior brief; list gaps and risky assumptions; do not rewrite the brief yet" is useful because it names the behavior you want.

Prefer checklists for review tasks

When you ask ChatGPT to review code, copy, or a support reply, ask for a checklist output: severity, finding, why it matters, suggested fix. Unstructured reviews wander.

Know when not to trust the output

Do not trust invented citations, medical or legal conclusions, or "exact" numbers the model produced without your data. Do not paste secrets into a chat you would not put in an email to a vendor. Better prompting does not cancel judgment.

Role-shaped prompts beat generic "AI tips"

A marketing brief and a code-review ask share the five-part scaffold. The context and constraints differ. That is why PromptDojo ships role paths instead of one generic "prompting course" identity.

Examples of how the ask changes:

If you are a marketer building pipelines you can run yourself, start with the marketer path. Developers who want out of "passenger in the IDE" mode: developer path. Ops folks automating repetitive department work: operations path.

A 20-minute practice block (do this today)

1. Pick one real task from your job (not a toy example).
2. Write a five-part prompt. Timebox: 5 minutes.
3. Run it. Score the draft: usable / half-usable / scrap.
4. Send one revision message that names what to keep, cut, and fix.
5. Save the final prompt + a note on what still failed.

Repeat tomorrow with a different task type. Skill grows from reps with feedback, which is why the dojo uses runnable steps instead of slide decks. The full map is on the curriculum page: 52 live chapters, 218 lessons, 1561 runnable steps in the browser.

Common mistakes to stop making

• Prompt stuffing. Ten personas and twenty rules in one blob. Prefer a tight first ask, then add rules when the draft shows the gap.
• Asking for originality without inputs. "Be creative" without audience, offer, or constraints produces generic copy.
• Letting the model invent proof. If you did not provide a number, do not publish the number it invented.
• One-shotting hard work. Complex jobs need stages: outline → draft → critique → revise.
• Skipping the human gate on customer-facing or compliance-sensitive text. Prompt skill includes knowing when a human must sign off.

How this connects to PromptDojo

Reading about prompts helps a little. Building the habit in a path helps more.

• Free start: Preview Chapter 00 in the company dojo. Short lessons, plain English in, a working builder loop out.
• Full map: Curriculum when you want the whole route.
• By job: Paths so marketers, developers, ops, PMs, and others do not take the same generic track.

You do not need a new personality to write better ChatGPT prompts. You need a clearer ask, honest constraints, a format you can edit, and the discipline to revise. Practice that on real work. The rest is volume.