Goal Loop Builder

Create a runnable contract for agent loops: a copy-ready command plus a durable Markdown file when useful. For /goal work, default to a real Markdown contract file plus a short one-line /goal command that references that file. Default to Chinese-first output with an English-compatible mirror unless the user asks otherwise.

Workflow

1. Classify the request.

   • Use a normal prompt for one-turn answers, small rewrites, or trivial shell output.
   • Use Codex /goal or Claude /goal when the task has one finish line and should keep advancing until evidence proves completion.
   • Use Claude /loop when the task should run on a timer, poll external state, babysit a PR/deploy, or produce recurring reports.
   • Use .claude/loop.md when the user wants a default prompt for bare /loop; it defines one default prompt, not a task list.
   • Use goal.md or <task-slug>.goal.md for /goal contracts; the Markdown file is the source of truth, and the /goal command only points to it.

2. Ask only for high-impact missing details.

   • Ask if runtime choice, target outcome, verification surface, write boundary, production access, credentials, or destructive authority is unclear.
   • Otherwise choose conservative defaults and state them briefly.