#agent

Instrument first, theorize second — logging with a stated plan for capturing the failure.

Walk the history between a known-good commit and main to find the culprit diff.

Audit finished work against the plan it was supposed to follow — gaps only, no style notes.

Close out a session with a why-focused commit and a PR that carries a test plan.

Update docs strictly from the diff, with no invented features.

Make the agent read and explain a subsystem before it touches a single line of it.

Find every caller a signature change would break, and the safe order to migrate them.

Migrate off a deprecated dependency one file at a time, stopping the moment tests fail.

A numbered plan with a verification step per phase, produced before any code is written.

Convert what went wrong into a one-line rule your agent config carries forward.

Have the agent read the ticket itself, find the relevant code, and propose a plan.

Point at a file you already like and have the agent bring another one in line with it.

Force a failing test that captures the bug before any fix is allowed to land.

Keep review scoped to what actually changed — correctness issues only.

Split a rough goal into tasks small enough that each is one focused session.

Delegate a read-only investigation so the findings come back without context bloat.

Turn a vague idea into a written spec by making the model interrogate you first.