AGENT SKILL

goal-meta-skill

Turn a long-running Codex request into one durable, evidence-verifiable goal. Use when the user asks to create, refine, review, or troubleshoot a Codex Goal or `/goal` instruction; do not use for ordinary one-turn prompts.

License 许可
MIT
最近更新
2026-11-03
Stars
★ 1
Forks
0

01 前言 Front Matter

name
goal-meta-skill
description
Turn a long-running Codex request into one durable, evidence-verifiable goal. Use when the user asks to create, refine, review, or troubleshoot a Codex Goal or `/goal` instruction; do not use for ordinary one-turn prompts.

02 资源文件 6

03 SKILL.md 正文

Goal Meta Skill

案例入口

先读 references/examples.md:其中给出正向案例、边界案例、失败恢复和可观察的验收证据;再按下文流程执行。

Design the smallest durable contract that still preserves the user's full intended outcome.

Establish the Goal

  1. Recover the real end state from the request and available project evidence. Write what must become true, not a list of activities.
  2. Identify the evidence that would prove the whole result. Match the evidence scope to the outcome scope; a narrow check cannot prove a broad claim.
  3. Preserve explicit requirements, named deliverables, constraints, permissions, and external-state dependencies. Do not silently turn a large goal into an easier milestone.
  4. Ask only when a missing choice materially changes product direction, external impact, cost, ownership, or risk. Otherwise state a conservative assumption in the goal.
  5. Produce one paste-ready /goal instruction in the user's language unless the user asks for analysis, alternatives, or another format.

Use references/goal-design.md when the request is vague, risky, multi-system, or difficult to verify.

Write an Executable Contract

The instruction should normally make these ideas clear in natural language:

  • Outcome: the observable state to achieve.
  • Evidence: tests, artifacts, runtime behavior, external records, or review criteria that prove it.
  • Invariants: behavior, data, interfaces, attribution, or user work that must remain intact.
  • Authority: what may be changed and which external, destructive, paid, public, or permission-changing actions still require authorization.
  • Persistence: how to use new evidence after failure without retrying the same assumption indefinitely.
  • Terminal states: what proves completion, and what constitutes a genuine human or external blocker.

These are design dimensions, not mandatory Codex field names. Do not claim that a fixed label set, bilingual mirror, numbered questionnaire, or exact prose template is required by the product.

Product Semantics

  • A Goal is for persistent work that benefits from repeated execution and verification. Keep simple answers and obvious one-step tasks as ordinary prompts.
  • Set a token budget only when the user explicitly requests one. Never invent a budget as a productivity tactic.
  • Do not replace an unfinished goal merely to rephrase it. If a goal is already active, refine it only through the product's supported flow or explain the conflict.
  • Do not encode new authorization inside the goal. A request to keep working does not authorize publication, deletion, payment, credential use, production changes, or other unrelated high-impact actions.
  • Treat schedules, indexing, approvals, third-party responses, and other external state as real dependencies. Never turn elapsed effort or repeated polling into evidence of success.

Read references/research-basis.md when checking product assumptions or updating this Skill for a new Codex release.

Check the Draft

For a saved draft, run:

python3 scripts/check_goal.py path/to/goal.txt

Use --format json for automation and --strict when warnings should fail CI. The checker detects structural omissions and risky contradictions; it cannot decide whether the proposed outcome is the right business decision.

Before returning the draft, confirm that:

  • every explicit deliverable remains in scope;
  • evidence proves the requested end state rather than a convenient subset;
  • placeholders and vague completion claims are gone;
  • risky actions retain their real authorization boundary;
  • completion and blocking are both grounded in authoritative state.

04 同仓库更多 Skills

siuserxiaowei
siuserxiaowei/xiaowei-ai-tools

x-article-draft-uploader

把本地 Markdown 转为经核验的 X Articles 新草稿。适用于需要保留首图封面、正文图片相对位置,并在独立 Playwright 会话中复用用户已授权 Chrome 登录态的任务;不负责公开发布。

★ 1
siuserxiaowei
siuserxiaowei/xiaowei-ai-tools

wecom-operations

通过用户已配置的官方 wecom-cli 管理企业微信文档、智能文档、待办、会议和日程。适合创建或查询这些云端对象、把本地 Markdown 转为智能文档,以及在明确确认后更新或取消对象;不发送消息、不控制桌面客户端。

★ 1
siuserxiaowei
siuserxiaowei/xiaowei-ai-tools

wecom-local-vault

Inspect authorized local WeCom databases on macOS, create private plaintext snapshots, and query or export contacts, conversations, and messages without modifying the source account store. Use when the user asks to parse 企业微信/WeCom local data, recover a private database snapshot, inspect local contacts or sessions, search chat history, or export a bounded conversation.

★ 1
siuserxiaowei
siuserxiaowei/xiaowei-ai-tools

wechat-reading

Read and analyze WeRead/微信读书 search results, book metadata, bookshelf entries, progress, notes, highlights, public reviews, recommendations, and reading statistics through the official read-only Agent Gateway. Use for WeRead account questions or exports; do not imply that the current Gateway can add, remove, upload, rate, like, or edit content.

★ 1
siuserxiaowei
siuserxiaowei/xiaowei-ai-tools

wechat-mp-batch-exporter

Build a bounded, private archive from known WeChat Official Account article URLs or an owner-authorized account-history export, with explicit reconciliation of article, publish-group, and platform-marked-original counts. Use for batch body capture, history normalization, or enhanced metrics/comments workflows. Do not use for a one-page summary or to automate the WeChat client.

★ 1
siuserxiaowei
siuserxiaowei/xiaowei-ai-tools

wechat-local-vault

在 macOS 本机建立并查询私有微信数据快照。适用于盘点数据库、离线校验已有密钥候选、创建全量或增量明文快照、查会话与联系人、按时间检索消息、导出指定聊天、查看朋友圈或收藏,以及准备群聊摘要素材。所有查询只读,默认隐藏内部账号标识,不操控微信界面,也不发送任何内容。

★ 1