goal-loop-builder

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.

1. Build the contract.

   • Outcome: one concrete result, not an activity.
   • Verification: commands, logs, screenshots, API responses, generated files, PR state, or other evidence.
   • Constraints: behavior, public API, data, style, security, and branch rules that must not change.
   • Boundaries: allowed write paths, forbidden paths, tools, environments, and external systems.
   • Iteration policy: small steps, rerun checks, inspect logs before retrying, change evidence source after repeated failure.
   • Stop/Pause: evidence that proves completion; blockers that require a human decision.
   • For /goal, write the full contract to a real Markdown file before final output. Use goal.md when safe; if it already exists, use <task-slug>.goal.md. In projectless workspaces, write to the available output directory or current workspace.
   • Keep the /goal command one line and under 350 characters. It should reference the Markdown file with @/absolute/path/to/file.md; never paste the full contract into the command.

2. Output in this order.

   • 推荐运行方式
   • 已生成 Markdown 文件 for /goal contracts
   • 可复制 /goal prompt or 可复制命令
   • 文件内容摘要 for /goal contracts
   • 默认选择理由
   • 可选调整 only when choices remain useful
   • English-compatible mirror

Reference Routing

• Read references/runtime-playbook.md when choosing between Codex /goal, Claude /goal, Claude /loop, .claude/loop.md, and goal.md.
• Read references/output-templates.md when drafting final commands or Markdown artifacts.
• Run python3 scripts/lint_goal_loop.py <file> for saved goal or loop artifacts before saying they are ready.

Quality Bar

Reject or revise output that:

• Leaves placeholders such as [goal], <path>, TODO, or 待补充.
• Uses vague verification such as "make sure it works" or "looks good enough".
• Allows broad writes like "edit anything" or "change whatever".
• Lacks a stop condition, pause condition, write boundary, or verification evidence.
• Puts a full /goal contract directly in the command instead of a referenced Markdown file.
• Produces a /goal command longer than 350 characters.
• Treats /loop as a completion-driven command. /loop is timer-driven; /goal is completion-driven.
• Treats .claude/loop.md as multiple scheduled tasks. It is one default prompt for bare /loop.

Examples

User: "Help me make an App"

Output: Recommend Codex or Claude /goal, create goal.md or app-mvp.goal.md, and return a short /goal Read @/absolute/path/to/file.md ... prompt.

User: "Have it check every 5 minutes whether the deployment is complete"

Output: Recommend Claude /loop 5m ..., include observation signals, reporting format, stop/escalation conditions, and optional .claude/loop.md.

User: "Fix the flaky test in the existing repository"

Output: Recommend /goal, create a file-backed goal contract, require discovery of test commands, isolated fix boundaries, regression evidence, and pause on missing repro or environment blockers.