Become an agentic engineer.
Ship production-ready code with AI agents — and prove it works.
Be first in line for cohort #1
Enrollment opens September 22, 2026 at the $497 founder price. Join and you get the dates the moment they're set, plus what you need for budget approval.
No spam. Unsubscribe anytime.
AI Code Craft
The agentic engineering cohort — workflow, guardrails, proof.
Fits a full-time job
5–8 h/week, fully async. Lessons drop Monday; you schedule them, not us.
Real code, not demos
Every exercise runs on an 8-year-old legacy codebase. The capstone runs on yours.
Proof, not vibes
You leave with a number: your test suite's trust score, before and after.
Yours to keep
Lessons, the repo, all five squad agents, recordings, transcripts. Forever.
“My current strategy is to not read any of the code written by my agents… I surround the agents with extreme constraints. Unit tests, gherkin tests, QA procedures, quality metrics, mutation testing, test coverage… they've had to run the gauntlet.”
This course is how you build the gauntlet.
Two ways to get this wrong.
Trust everything
The agent ships 40 green tests. You merge. Karpathy calls these mistakes “subtle conceptual errors a slightly sloppy, hasty junior dev might do.” [source]
Trust nothing
You read every line. Now you're the bottleneck.
Agent PRs wait 5.3× longer for review — LinearB, 2026The middle path is verification: machines prove the code, you judge the proof. It's a system — and you can build it in two weeks.
From reading diffs to judging proof.
| Today | After the cohort |
|---|---|
| Review every diff line-by-line | Verify at the boundaries that matter |
| Hope the tests catch it | Prove it before merge |
| 40 green tests, unknown value | A trust score per module |
| Courage-based autonomy | Autonomy granted by the gate |
Steer, then verify.
Week 1 puts your hands on the controls. Week 2 builds the proof that lets you let go. Every day ends with something shipped.
RampGet readyget ready
Unlocks the minute you enroll. One evening, nobody arrives stuck.
lessons +
- Start Here — the map, pacing, office hours, Discord
- Cockpit Basics — install, sessions, context controls, rewind
- Models, Plans & Token Budget — what this course costs to run
- The Engineering Taxonomy — harness, loop, context, memory, prompt
- Engine vs Cockpit — terminal, Slack, Discord are steering wheels
- Permissions & Sandboxing — an agent that cannot hurt you
- Repo Setup — guided, with fix-it skills in the repo
D1 · W1Read the Codebase Like an Agent
Old code, made legible.
lessons +
- Meet the repo — 8 years of rot, one god module
- How agents actually read code — the smart zone + the four context failure modes
- Session discipline — one task, fresh context, handoff not compaction
- Agent archaeology — subagents map the god module
- Ship it — AGENTS.md v1 + one merged agent change
You ship: A working AGENTS.md for a codebase you didn't write, plus merged code.
D2 · W1Safety Net First
The agent does not edit code that has no net under it.
lessons +
- Green means nothing yet — the planted bug your tests miss
- Characterization tests — the agent writes the net that will catch the agent
- Seams — make it testable without changing behavior
- CI as a tripwire — the net goes into CI today (gate v0)
- Ship it — first agent refactor of the god module, provably harmless
You ship: A pinned god module + one agent refactor proven safe.
D3 · W1Steering: Skills & Specs
Stop prompting. Start managing.
lessons +
- Skills, not prompts — progressive disclosure, the harness ratchet
- The gotchas loop — every agent mistake makes the repo smarter
- Grill before you build — the agent interviews you
- Slice vertically — tracer bullets through existing layers
- When a skill isn't enough — build an agent-shaped MCP server
- Ship it — skill + grilled spec + sliced tickets, first ticket merged
You ship: A steering kit: one custom skill, one agent-shaped MCP for the repo, one merged ticket.
D4 · W2The Trust Engine
Measure whether green means anything.
lessons +
- The agent's tests are lying to you — 40 green tests, several hollow
- The Mutant Hunter — mutation testing, survivors, the trust score
- The Property Prover — invariants your example tests miss
- TDD's survivor — test-first bug fixing with fresh-context runs
- Ship it — the Gate List, wired into CI: done-skill → pre-commit → thresholds
You ship: A CI gate that fails when the agent's tests are hollow.
D5 · W2Review at Scale
One PR, five verdicts.
lessons +
- The 5.3× problem — reviewing became the bottleneck
- The Adversary — fresh-context review vs spec and standards
- The Runtime Witness — the agent that actually runs your app
- Trust tiers & diff budgets — attention where it matters, costs capped
- Ship it — the full squad wired into every PR
- Greenfield: the easy case
You ship: The Verification Squad running on every pull request.
D6 · W2Delegation at Scale
Autonomy granted by the trust score, not courage.
lessons +
- The autonomy ladder — pair → parallel → unattended (and the rungs beyond)
- One dev, many agents — three tickets, three worktrees, live
- The night shift — unattended, shown honestly, failures included
- The cockpit swap — the same engine, driven from Discord
- Capstone briefing
You ship: One backlog ticket completed by agents while you reviewed verdicts.
CapCapstone — Your Repo, Verified
The whole pipeline on your own codebase. Results, not code.
details +
- Safety net on one module of YOUR repo
- AGENTS.md + one custom skill for YOUR repo
- A verification gate in YOUR CI
- One agent task shipped through that gate
- The number: your mutation-score delta, before → after
Squad-reviewed submission; the best dozen get a live teardown in the final office hours.
Five agents. Each with one job.
They're markdown skills, not magic — one is open source, so you can read exactly what you're buying. You build all five. You keep them forever.
Writes characterization tests that freeze current behavior before anything changes.
Runs Stryker, reads the survivors, reports the trust score. Open source — run it today.
Finds the invariants your example tests miss, and hands the agent properties as spec.
Reviews the diff in a fresh context, against the spec and your standards. Issues first, no praise, merge verdict.
Runs the app, clicks the buttons, reads the console. Tests can pass while the app is broken; this one notices.
No student results yet. Inspect the work instead.
Cohort #1 hasn't run — that's what the founder price is for. Everything below ships before enrollment opens:
The practice repo, public
Eight years of realistic rot. Every exercise runs here. Clone it.
ships before Sep 22One squad agent, open source
The Mutant Hunter — run it on your repo this afternoon.
ships before Sep 22A full sample lesson, free
Lesson 4.2, complete. The product experience, not a trailer.
ships before Sep 22One unedited demo
The squad catches a real bug on camera. No cuts.
ships before Sep 22And one commitment you can hold me to: I'll publish every cohort-1 capstone's mutation-score delta — including the misses.
One price. One window.
Cohort-1 founder price. Enrollment runs September 22, 2026 → October 2, 2026. Cohort #2 lists at $995 — early buyers are never repriced.
Team: 3+ seats: 10% off, one invoice. VAT invoice at checkout. Letter for your boss →
- ~40 tight lessons, video + written guide + quizzes
- The legacy practice repo, checkpointed
- All five Verification Squad agents
- 6 live office hours, two time zones, recorded
- Cohort Discord + capstone review & showcase
- Lifetime access. No subscription
This is a presale. The cohort runs Nov 9–20, 2026; the course is built after enrollment confirms demand. Two protections, in writing: a 30-day no-questions refund, and if fewer than 20 people enroll, it doesn't run — everyone is refunded 100%, automatically.
The questions you should ask.
The course isn't built yet. Why pay now?
That's what the 50% founder discount buys: you carry the early risk, I price it accordingly. The curriculum, repo, sample lesson and demo are finished and inspectable before enrollment opens. The refund terms remove the rest.
You sunset your last course. Why would this one live?
The TDD course earned $25K, had zero refunds, and was still selling when I killed it — because AI changed the practice and I don't sell what I no longer believe. That standard is why this course is demand-gated and refund-backed. The thinking didn't die; it moved up a level — from testing code to verifying agents. This course is where it went.
Who is this for — and not for?
For engineers who run Claude Code, Codex, or Cursor on real work, ship to production, and don't fully trust the output. Comfortable in TypeScript; techniques stay language-independent, with C#, Java, and Python equivalents named. Not for beginners, and not a prompt-tricks course.
Honest option: my free videos cover the ideas. If that's enough — don't buy. The cohort is the reps, the repo, and the proof.
Mutation testing is slow and noisy. I've tried it.
True — if you run it naively. The course teaches the practical version: scoped runs on the diff, survivors triaged by an agent, deltas instead of absolutes. That's what makes it a gate instead of a science project.
Why not Matt Pocock's AI Hero cohort?
It's good — for a different job. AI Hero teaches breadth: steering, planning, AFK workflows. If you want that, take it. AI Code Craft makes one narrow claim: proving agent-written code correct on real, aging codebases — with a number. Verification is a garnish there. It's the whole meal here.
Which tools do I need?
An agent harness (lessons demo Claude Code; Codex/Cursor equivalents stated), Node 20+, a Claude Pro plan or equivalent. The techniques are tool- and language-independent.
How much time per week?
5–8 h/week. All lessons drop Monday — pace yourself. Office hours run in two time-zone slots and everything is recorded.
Can't use company code for the capstone?
Same checklist on a spare module of the practice repo. You never submit code anyway — results only.
Where are the testimonials?
There are none — cohort #1 hasn't run. Inspect the proof section instead. After cohort 1, this section shows every capstone delta, including the bad ones.
Enrollment opens September 22, 2026.
The waitlist hears first — dates, founder price, and everything you need for budget approval.
No spam. Unsubscribe anytime.