Where this course goes (and where you fit) — step 2 of 9
Who's scarce in 2026
There are two skills that hiring managers are paying for right now. One is writing prompts that don't waste tokens or produce garbage. The other is writing code that doesn't fall over when the model hallucinates inside it. Almost everyone can do one. Almost nobody can do both.
That intersection — prompt fluent AND code fluent — is the niche this course is aimed at. It is small. That's why it pays.
Every AI feature shipping today needs someone who reads outputs like a paralegal reads a contract. That person is rare. That person is you, plus six weeks.
The role spectrum
The intersection isn't a single job. It's a spectrum. Different chapters of this course serve different end-states on it.
- Prompt engineer. The most accessible role. You write, version, and grade prompts. Less code, more language. The chapters that matter most: ch01-04 (basic Python so you can read code), ch13 (LLM APIs), ch19 (prompting), ch21 (evals). AI-engineering roles span a wide range: junior in-house roles cluster lower, senior or agency roles cluster higher. We won't quote a specific salary band — the market has been moving every quarter — but most of the public hiring data (LinkedIn Workforce Reports, Levels.fyi) shows the AI-skilled premium is real for senior engineers and unclear at the junior level. Treat any specific dollar number you read from a course or a recruiter as an aspirational data point, not a guarantee.
- Eval engineer. The role that grew out of "the prompts are good but how do we know?" You write graders, build test sets, run regression tests against new model versions. The chapters that matter most: ch11 (classes), ch14 (structured output), ch20 (agent traces), ch21 (evals). This is the most underrated role in the market right now.
- Agent builder. You wire models into loops, tools, and retrieval. You build the thing that actually runs. The chapters that matter most: all of ch12-25. The bulk of the course aims here.
- Harness engineer. The advanced role. You build the scaffolding that makes agents reliable in production — the retry logic, the observability, the cost guards, the failover. Chapters: ch22-30. Almost nobody has done this for two years yet, because the role didn't exist two years ago. Top of the market.
You do not need to pick one now. Most people end up sliding along the spectrum as they learn what they're good at. The course is built so that finishing ch01-21 puts you in prompt engineer or eval engineer range, ch01-25 puts you in agent builder range, and ch01-30 puts you in harness engineer range. Each tier is a real job.
The "differently technical" advantage
If you came from copywriting, support, paralegal work, or design — you already do something that most engineers can't:
- You can read a paragraph and tell whether it's lying. Engineers often can't. They check syntax, not semantics.
- You can write a brief that another human can actually execute. Engineers often can't. They write specs for compilers, not people.
- You have tolerance for ambiguous, judgment-driven, partly-subjective work. Engineers often don't. They want type signatures.
Those three skills are exactly what eval engineering and prompt engineering are. You are not catching up to engineers. You are adding their layer to a skill set they don't have.
By chapter 30 you won't be "using AI." You'll be the person other engineers ask how to use it.
If you came from engineering — you have the inverse advantage. You can already read tracebacks, ship to production, reason about concurrency. The thing you're adding is the language layer: the discipline of writing prompts that don't waste the model and reading outputs like a paralegal reads a contract. That's why your path through the course skips different chapters than the copywriter's path does. We'll map both paths in the next two lessons.
What this means concretely
The market does not pay for "I learned Python." It pays for "I can ship the thing that does the job you used to pay a junior to do."
Every chapter of this course is calibrated to that bar. By the time you finish, you should be able to point at a workflow at a company and say: "I could rebuild that as an agent. Here's what it would cost. Here's how I'd grade whether it works."
That sentence is what the intersection sounds like out loud. We're going to get you there.
What these personas leave out
The next step asks you to route four learner profiles — Maya, Devon, Priya, and Jules — through this course based on their background and weekly hours. The personas are deliberately simple. They have to be, because a multiple-choice question can't model a real life.
Real displaced workers come with more variables than this exercise can model:
- Childcare and caregiving load. A 38-year-old single parent retraining at night, after the kids are down, has a different curve than a 28-year-old with no dependents. The "10 hours a week" looks the same on paper and is not the same in lived time.
- Visa status. A displaced engineer on H1-B has 60 days to find a sponsoring employer or leave the country. The calculus is not "reskill at your own pace." It is "ship something employable in eight weeks or leave the life you built." This course is too slow to be the entire answer for that worker. It can be part of one.
- Non-remote schedules. A paralegal who works 9-6 in person at a firm and is retraining on the side can't take the "lunch hour" study slot the personas implicitly assume. Their study time is evenings, weekends, and the 20 minutes on the train.
- Financial runway. The student who is one missed mortgage payment from foreclosure cannot afford a four-month investment without a credible bridge income. The course gives you a skill in 3-6 months. It does not pay rent. If the runway is shorter than the timeline, the right plan is parallel — take this on the side of whatever pays rent today.
- Mental-health load. Grief over a lost job is not resolved by finishing lesson 02. Anger, shame, fear, and the slow erosion of identity that comes with being laid off are real variables in how much capacity you have on any given week. The plan that ignores this is the plan that breaks in week three.
If you don't match the personas exactly — and most readers won't — pick the persona closest to your hours-per-week, then adjust your own plan for the variables this course can't model for you. The course can give you a skill. It cannot give you a different life. The honest version of "you can do this" is "you can do this, on top of everything else you are already carrying, if you build the plan around what you are actually carrying instead of around the persona this exercise asks you to pick."
Where this course goes (and where you fit) — step 2 of 9
Who's scarce in 2026
There are two skills that hiring managers are paying for right now. One is writing prompts that don't waste tokens or produce garbage. The other is writing code that doesn't fall over when the model hallucinates inside it. Almost everyone can do one. Almost nobody can do both.
That intersection — prompt fluent AND code fluent — is the niche this course is aimed at. It is small. That's why it pays.
Every AI feature shipping today needs someone who reads outputs like a paralegal reads a contract. That person is rare. That person is you, plus six weeks.
The role spectrum
The intersection isn't a single job. It's a spectrum. Different chapters of this course serve different end-states on it.
- Prompt engineer. The most accessible role. You write, version, and grade prompts. Less code, more language. The chapters that matter most: ch01-04 (basic Python so you can read code), ch13 (LLM APIs), ch19 (prompting), ch21 (evals). AI-engineering roles span a wide range: junior in-house roles cluster lower, senior or agency roles cluster higher. We won't quote a specific salary band — the market has been moving every quarter — but most of the public hiring data (LinkedIn Workforce Reports, Levels.fyi) shows the AI-skilled premium is real for senior engineers and unclear at the junior level. Treat any specific dollar number you read from a course or a recruiter as an aspirational data point, not a guarantee.
- Eval engineer. The role that grew out of "the prompts are good but how do we know?" You write graders, build test sets, run regression tests against new model versions. The chapters that matter most: ch11 (classes), ch14 (structured output), ch20 (agent traces), ch21 (evals). This is the most underrated role in the market right now.
- Agent builder. You wire models into loops, tools, and retrieval. You build the thing that actually runs. The chapters that matter most: all of ch12-25. The bulk of the course aims here.
- Harness engineer. The advanced role. You build the scaffolding that makes agents reliable in production — the retry logic, the observability, the cost guards, the failover. Chapters: ch22-30. Almost nobody has done this for two years yet, because the role didn't exist two years ago. Top of the market.
You do not need to pick one now. Most people end up sliding along the spectrum as they learn what they're good at. The course is built so that finishing ch01-21 puts you in prompt engineer or eval engineer range, ch01-25 puts you in agent builder range, and ch01-30 puts you in harness engineer range. Each tier is a real job.
The "differently technical" advantage
If you came from copywriting, support, paralegal work, or design — you already do something that most engineers can't:
- You can read a paragraph and tell whether it's lying. Engineers often can't. They check syntax, not semantics.
- You can write a brief that another human can actually execute. Engineers often can't. They write specs for compilers, not people.
- You have tolerance for ambiguous, judgment-driven, partly-subjective work. Engineers often don't. They want type signatures.
Those three skills are exactly what eval engineering and prompt engineering are. You are not catching up to engineers. You are adding their layer to a skill set they don't have.
By chapter 30 you won't be "using AI." You'll be the person other engineers ask how to use it.
If you came from engineering — you have the inverse advantage. You can already read tracebacks, ship to production, reason about concurrency. The thing you're adding is the language layer: the discipline of writing prompts that don't waste the model and reading outputs like a paralegal reads a contract. That's why your path through the course skips different chapters than the copywriter's path does. We'll map both paths in the next two lessons.
What this means concretely
The market does not pay for "I learned Python." It pays for "I can ship the thing that does the job you used to pay a junior to do."
Every chapter of this course is calibrated to that bar. By the time you finish, you should be able to point at a workflow at a company and say: "I could rebuild that as an agent. Here's what it would cost. Here's how I'd grade whether it works."
That sentence is what the intersection sounds like out loud. We're going to get you there.
What these personas leave out
The next step asks you to route four learner profiles — Maya, Devon, Priya, and Jules — through this course based on their background and weekly hours. The personas are deliberately simple. They have to be, because a multiple-choice question can't model a real life.
Real displaced workers come with more variables than this exercise can model:
- Childcare and caregiving load. A 38-year-old single parent retraining at night, after the kids are down, has a different curve than a 28-year-old with no dependents. The "10 hours a week" looks the same on paper and is not the same in lived time.
- Visa status. A displaced engineer on H1-B has 60 days to find a sponsoring employer or leave the country. The calculus is not "reskill at your own pace." It is "ship something employable in eight weeks or leave the life you built." This course is too slow to be the entire answer for that worker. It can be part of one.
- Non-remote schedules. A paralegal who works 9-6 in person at a firm and is retraining on the side can't take the "lunch hour" study slot the personas implicitly assume. Their study time is evenings, weekends, and the 20 minutes on the train.
- Financial runway. The student who is one missed mortgage payment from foreclosure cannot afford a four-month investment without a credible bridge income. The course gives you a skill in 3-6 months. It does not pay rent. If the runway is shorter than the timeline, the right plan is parallel — take this on the side of whatever pays rent today.
- Mental-health load. Grief over a lost job is not resolved by finishing lesson 02. Anger, shame, fear, and the slow erosion of identity that comes with being laid off are real variables in how much capacity you have on any given week. The plan that ignores this is the plan that breaks in week three.
If you don't match the personas exactly — and most readers won't — pick the persona closest to your hours-per-week, then adjust your own plan for the variables this course can't model for you. The course can give you a skill. It cannot give you a different life. The honest version of "you can do this" is "you can do this, on top of everything else you are already carrying, if you build the plan around what you are actually carrying instead of around the persona this exercise asks you to pick."