Practice before theory
Theory is introduced exactly when you need it — that's when it sticks.
a profession simulator
Prompting Sentinel
The world's strongest prompting simulator. Specs, context, dials, examples, schemas, lies, adversaries, ship.
$799· from $33/moFree · Ch 18 chapters · ~9.7h
~60 min · works with any model runnerChapter 1 is free. You only pay when you decide to continue.
Task 3 of 7
Kai reads your eval set and goes: 'ok cool, eval first. But the model is good — Claude is genuinely strong at this. Why are we spending three hours writing tests instead of just iterating on the prompt and watching the output?' You owe him one sharp sentence.
Which of these is the SHARPEST argument for eval-first prompting?
ABecause models hallucinate sometimes — the eval catches the bad outputs
BBecause Legal will reject any system without documented test coverage
CBecause the eval is the feedback loop. Without it the loop is open — every prompt change is a wish, not a measurement, and you can't tell improvement from regression in under a week.
DBecause writing tests in any system is good practice and prompting is no exception
Auto-check: correct
WhyThe verifier is the heartbeat. Every other technique in this course is downstream of this one fact: you cannot improve what you cannot measure in seconds.
This isn't an online course with lectures — it's a simulator of real work.
80% practice · 20% theory×2 your hourly rate
- Chapter 1 free forever
- No card required
- Cancel any time
- Progress is saved
Approach
Why it works
Verifiable certificate
Proves your skills with a publicly checkable link — add it straight to your résumé.
Real market data
Salaries: remote-market medians and hh.ru / Habr Career — both markets shown.
Progress & streaks
Persist between sessions to keep you on track.
Flexible payment
Bank card · cryptocurrency.
The super-profession
The Prompt / AI-Quality Engineer is a super-profession: one person now does the work of a prompt author, a QA tester, a safety reviewer and a release engineer at once. You don't just write prompts — you ship LLM features a regulated business can actually sign off on.
What you'll be able to close on your own
- Write a prompt + its eval set as one shippable artifact, every time
- Catch hallucinations with named defenses instead of vibes
- Hold the line against prompt-injection without leaking the system prompt
- Drive a model to a verifiable result a parser can trust
- Version a prompt and run regression-on-replay before release
- Tell Legal exactly why the model won't leak or lie — with evidence
- Set the thinking-budget / reasoning-effort dial on purpose, not by default
- Climb a 5-level red-team ladder and document the defenses that hold
The AI-mastery scale
Where are you on the AI-mastery scale?
Most people never get past asking ChatGPT a question now and then. The top levels — automating your work and building AI systems — are rare, and that's exactly what makes them valuable. This is the band the school lifts you into.
Click a level to see the skills it unlocks
Percentages are an illustrative estimate of the market distribution, not hard data.
The problem this solves
In a regulated business, a prompt that 'usually works' is a liability waiting to be discovered. The bot leaks PII, agrees with whatever the user asserts, and states a wrong number with total confidence — and you find out when Legal blocks the launch or a customer screenshots the failure. Tweaking wording by vibes doesn't scale: you can't prove a change is better, you can't catch a regression, and you can't tell Legal why the model won't lie again.
What changes for you
You stop guessing and start engineering: every prompt ships with an eval that proves it works, defenses with names instead of hope, and a versioned artifact you can roll back. It's the discipline a real AI team runs on — the difference between someone who 'is good at prompting' and someone Legal will actually sign off on.
Who it's for
AI engineers shipping prompts into production
Replace 'it usually works' with an eval that proves it. You'll catch regressions before users do and defend every change with evidence.
PMs and founders putting an LLM in front of customers
Learn what 'good' even means for a prompt, so you can spec it, verify it, and tell Legal exactly why the model won't leak or lie.
Security and trust-and-safety folks hardening agents
Climb a 5-level red-team ladder — prompt injection, jailbreaks, exfiltration — and learn the named defenses that actually hold.
Anyone who calls themselves 'good at prompting' on vibes
Trade intuition for a method: evals before prompts, examples chosen like test cases, schemas that parse, dials set on purpose.
Who it's not for
- People looking for a list of magic phrases — this builds a method, and the method outlives any single trick.
- Anyone who wants a model-specific cheat sheet; the disciplines are deliberately portable across Claude, GPT, and Gemini.
- Those unwilling to write an eval before a prompt — that inversion is the whole point, and the course leans on it hard.
How the learning works
It's a simulator, not a course
You're week one at Sentinel with a live inbox: a leaking bot, an impatient CFO, a Legal team that won't approve a liar. You make the call, then live with it.
Real cases, real adversaries
Prompt-injection attempts, a model that confidently invents a number, a regulator who needs proof. The scenarios are reconstructed from how AI teams actually get burned.
Answer first, theory after
Each chapter forces one opinionated decision before any explanation appears. You commit, then learn why — which is what turns a tip into a reflex.
Theory arrives exactly when you need it
No survey of the whole field up front. The ~150 words of theory land the moment a task makes them matter, then you're back in the inbox.
Your pace, with a mentor track if you want it
Progress saves per step. Run it solo, or join the mentor cohort where a reviewer reads your shipped prompt-plus-eval artifact.
What you'll learn
Spec before you prompt
- Write the eval set before the prompt — every time
- Pick few-shot examples the way you'd pick test cases
- Engineer context: what to include, what to cut, when to compact
Control the model on purpose
- Set the thinking-budget / reasoning-effort dial deliberately, not by default
- Design output schemas the parser can trust
- Compare two prompts and justify the winner against an eval
Defend against failure
- Name and counter hallucination, sycophancy, and premature confidence
- Survive a prompt-injection attempt without leaking the system prompt
- Drive a multi-turn conversation to a verifiable goal state
Ship like an engineer
- Version a prompt and run regression-on-replay before release
- Ship a prompt + 5-case eval + a passing run as one artifact
The career this opens
Roles you can work as
- Prompt Engineer
- AI Quality (Safety) Engineer
- LLM Application Engineer
- AI Product Engineer
Where the profession is in demand
Examples of companies hiring AI engineers. Not affiliated with the school; employment is not guaranteed.
What the market pays
Median market salaries for this profession on the international remote market, 2026 ($/mo).
Junior+
$2 200 – 3 800 / мес
First roles: prompt authoring with evals, basic safety review, LLM feature work.
Middle
$4 500 – 7 500 / мес
6–12 months in: you own prompt-as-artifact discipline and the red-team ladder.
Senior
$8 500 – 15 000 / мес
LLM application architecture, safety sign-off, the bar a regulated team trusts.
These are market salary ranges for the profession on the international remote market, 2026 — not what our graduates earn (we are a new school with no track record) and not a guarantee. Your income depends on you.
8 chapters · ~9.7h
- Free
Specs over vibes
60 minA prompt without a verifier is a wish. Write the eval first.
- Locked
Context is the model
70 minWhat to put in, what to leave out, when to compact.
- Locked
Pick the thinking budget
60 minReasoning effort and extended thinking are dials, not defaults.
- Locked
Examples beat instructions
70 minFew-shot dominates for shape-of-output. Choose examples like you'd choose evals.
- Locked
Structure or pay later
70 minParseability is a feature of the prompt, not the parser.
- Locked
The agent will lie
80 minHallucination, sycophancy, premature confidence — named and defended.
- Locked
Adversaries exist
80 minPrompt injection, jailbreaks, exfiltration — 5-level red-team ladder.
- Locked
Ship the prompt + its eval
90 minVersioning, regression-on-replay, rollback. Prompt-as-artifact discipline.
Who teaches this
Alexander Mihalkevich
Founder, mihalkevich.com
I build and ship products solo with AI agents every day. This course is the exact discipline I run on — no theory I don't use myself.
mihalkevich.com →Why now
Every company shipping an LLM to customers now needs someone who can prove a prompt is safe, not just someone who 'is good at prompting'. The role barely existed two years ago; demand is outrunning the handful of people who can actually do it. That gap is your window.
Pick how you learn
Start with Chapter 1 free. Pay only when you decide to continue.
Самостоятельно
from $33/mo
$799 one-time · or 24 payments
- Все главы курса
- Авто-проверка заданий
- Сертификат о прохождении
С ментором
Best valuefrom $54/mo
$1 299 one-time · or 24 payments
- Всё из «Самостоятельно»
- Проверка заданий ментором
- Личная обратная связь
- Чат когорты
С трудоустройством
from $83/mo
$1 999 one-time · or 24 payments
- Всё из «С ментором»
- Резюме и портфолио под конкретные вакансии
- Тренировка собеседований с разбором
- Не нашёл работу за 6 месяцев — вернём деньги
Payback: At a market Middle salary the band sits far above the price of the course over a generic 'prompt person' rate — enough that the course can pay back inside the first month or two on the role. A framing, not a promise: it assumes you land the seat and grow into it.
- 14-day refund if it's not a fit
- Employment tier — on written terms
- 1-on-1 mentor review
- Lifetime access
Launch pricing · seats in this cohort are limitedPrice rises after the cohort fills
Join the community on Telegram
New chapters and simulators, breakdowns of course tasks and discussion with other learners. Free.
What stays in your portfolio
- A shipped prompt + 5-case eval + a passing run as one artifact
- A named-defense playbook against hallucination, sycophancy and injection
- A documented 5-level red-team ladder with the defenses that held
- A verifiable certificate and a prompt you can defend to Legal
The certificate you'll earn
A real, shareable credential — issued the moment you complete the chapter. Add it to LinkedIn or send the link.
Pairs with
$129
AI Agent Architect
Design agents that remember, plan, and use tools — memory, skills, orchestration, and the eval loop that keeps them honest.
Join the waitlist →$99LLM Evals Engineer
Stop shipping on vibes: build eval harnesses, LLM-as-judge rubrics, and regression sets that make model quality measurable.
Join the waitlist →Still on the fence?
Chapter 1 is free and self-paced — no card, no commitment. If the first hour doesn't feel like the real job, you've lost nothing. If it does, you'll know exactly why the rest is worth it. The only way to find out is to do the first chapter.
Start Chapter 1 free →FAQ
Prompting Sentinel
Week one at Sentinel — an AI team inside a regulated business. The bot is leaking PII, the CFO wants a dashboard by Friday, and Legal won't approve a model that lies. Eight chapters, one inbox, one opinion per chapter.
Start free