The memory file
CLAUDE.md and AGENTS.md that actually get read and obeyed. What belongs in one, what poisons it, and why most are ignored by the model within two turns.
Rung 02 · Build
Stop prompting.
Start building a harness.
Almost everyone using an AI coding tool is still typing wishes into a box and hoping. The people getting real output built something around the model — memory, tools, a verification loop. That thing is the harness, and it is teachable in an afternoon.
$ claude "make the payment tests pass"
→ run: pnpm vitest run payments
✗ 2 failing — refund rounds the wrong way
→ patch, re-run
✓ 14 passing — stopping because the check passed
£89
Self-paced · lifetime access · free updates
What's in it
Six modules, each ending in something you use tomorrow
Plan before execute
Getting the plan out of the model before it touches a file — and how to review a plan so it catches the wrong approach while it is still cheap.
Subagents in parallel
Splitting work across agents that do not share state, and knowing when a barrier is genuinely needed versus when it is just wasting wall-clock.
Real tools, via MCP
Giving an assistant your actual systems instead of letting it guess. Scoping what it may touch, and what should never be in reach.
Verify, then loop
The single highest-value habit: a script that decides pass or fail, and a loop that runs until it passes. This is what separates a demo from a result.
Evals, not vibes
A ten-row table that turns "it feels better" into a number. How to build one in an afternoon and why it changes every argument you have about a model.
Who it's for
- Working developers whose AI output is inconsistent and who can't say why
- Teams that adopted an assistant and saw no measurable change
- Technical founders shipping alone who need the leverage to be real
Who it isn't for
- Anyone wanting a prompt pack — there isn't one, and they don't work
- Complete beginners: you should be able to read and run your own test suite
- Anyone hoping to remove review from the loop. The gate is the point
Being built now
First cohort, honestly
The course is in production and not yet on sale. The first run will be priced at £89 and capped deliberately small, because the first cohort's feedback is worth more than the revenue.
Rather than collect emails for something that doesn't exist yet, the library is the honest place to start — it's free, it's live, and it's the same voice.
No payment is taken on this page. When the cohort opens, it opens properly.