Goal prompt

goal-prompt

Produce a ready-to-paste /goal ... command from whatever the user is trying to accomplish.

Why the shape matters

/goal runs Claude autonomously until a separate fast-model evaluator decides, after a turn, that the condition is met. The evaluator does NOT run commands or read files itself — it only reads Claude's output. So the completion condition must be demonstrable by Claude's own output, never by hidden side effects.

A good goal has three parts

1. Measurable end state — one concrete finish line: a test/exit code, a file that must exist, a count, an empty queue, named sections present.
2. Stated proof — exactly how Claude demonstrates it: the command to run and its expected result, or the grep/check whose output shows done. Phrase it as "Prove it by showing X."
3. Constraints that must not drift — what stays unchanged on the way there: files not to touch, framing to keep, no network/prod, don't modify tests.

How to write it

• One sentence of objective, then Done when: <end state + proof>, then Constraints that must not change: <list>.
• If the full spec is long, point to a plan/doc file (e.g. a path under ~/.claude/plans/ or docs/) and keep the goal itself scannable.
• Make the proof something the transcript can show: prefer command exits 0 + a summary line, or grep for markers, over vague "it works".
• Translate conditions the evaluator can't see ("the UI looks good") into an observable check.
• Keep constraints tight enough to stop scope creep, not so rigid they block the obvious path.

Output

Give the user a single fenced block starting with /goal, then 2-3 lines explaining the end state, the proof, and why it's demonstrable. Nothing else.

Example

/goal Add a --json flag to the export CLI per docs/export-json.md. Done when: `pytest tests/test_export.py -q` exits 0 and `python -m app.export --json` prints valid JSON whose top-level keys include "rows" and "meta". Prove it by showing the pytest summary and the piped `... --json | jq keys` output. Constraints that must not change: only edit app/export.py and add tests/test_export.py; do not alter the existing CSV output path; no network.

Its finish line is a passing test plus a schema check, both visible in Claude's transcript; the constraints pin the blast radius so the autonomous run can't wander.