Step 1 / 5 · 15 min
Give the agent enough context to stay narrow and testable.Write the contract
- Describe user behaviour first.
- Name relevant files and invariants.
- Write non-goals and verification before prompting.
Build · Core skill lab
Agent Brief → Reviewed Diff
Turn a coding assistant from an improviser into a bounded implementer whose work you can explain and verify.
- Decision
- Is this agent-produced change small, understood, and safe enough to merge?
- Timebox
- One 60–120 minute change
- Output
- Agent change review
Bring to the bench
- A working repository
- A narrow user-visible change
- A clean verification baseline
Your work saves in this browser.
Field tools
Use the instrument, not a blank page.
Copy these into your interview, agent, review, or working document. They are specific to this repetition.
prompt
Bounded coding brief
Fill this before asking for implementation.
GOAL [user-visible behaviour] CONTEXT [relevant files, patterns, commands] ACCEPTANCE [observable cases] BOUNDARIES [non-goals, invariants, user-owned changes] VERIFICATION [commands + manual checks] FIRST RESPONSE Inspect only what is needed. Return a plan, risks, likely files, and questions. Do not edit yet.
checklist
Six-pass diff review
Review in passes so a plausible happy path does not end the inspection.
1. Behaviour: does it satisfy each acceptance case? 2. Boundaries: unrelated changes or hidden scope? 3. Failure: errors, loading, empty, retry, cancellation? 4. Trust: auth, secrets, validation, destructive actions? 5. Access: keyboard, focus, labels, responsive states? 6. Proof: tests and manual checks match the changed risk?
Calibrate judgment
Compare the evidence, not the polish.
A reviewed agent change
An assistant adds a clear-filters action to an empty search state.
Plan review removed a proposed store rewrite. Diff review found focus returned to the page body after clearing. The change stayed unmerged until the keyboard path and focused tests passed. The acceptance template now requires a focus destination.
Why it works: Human judgment bounds scope, finds a behavioural defect, reproduces evidence, and improves future briefs.
Delegation without review
Asked the agent to improve search UX. It changed nine files and said tests pass. The diff looks reasonable, so merged.
Why it fails: The goal and boundaries are ambiguous, the diff is not understood, and verification is delegated to the same system that wrote the change.
Review → revise → repeat
The artifact is the beginning of the rep.
Check only standards your current artifact actually meets. Then record one consequential revision before exporting it.