goals

For substantial work I write a goal rather than a chat message: the real outcome in one line, the exact files and surfaces it touches, hard constraints, how each change gets verified, and the artifact I want back. A well-written goal runs start to finish without me.

Half-stated goals produce half-right work that needs a second round anyway. Writing the outcome, constraints, and bar once, up front, is cheaper than discovering them through corrections — and it makes the run repeatable.

The goal text is the contract — it names what done means so the session can hold itself to it. One-shot goals cover a bounded task; longer ones decompose into parallel workstreams with the expensive model reviewing rather than typing. I keep a record of which goal shapes worked, and consult it before writing the next one.

The whole method is downloadable and adaptable — method/steps/goal-prompt.md inside the zip, with the structure I use: outcome, surfaces, constraints, verification bar, handoff artifact.