goal-meta-skill

name	goal-meta-skill
description	Turn vague or complex agent tasks into strong `/goal` commands for Claude Code and Codex — outcome, verification, constraints, boundaries, iteration policy, and stop conditions. Use for Codex/Claude Code goal instructions, Goal 指令, 目标指令, `/goal` prompts, 中文 Goal 模板, plan-to-goal interviews, or bounded agent work definitions.
version	0.2.0
category	developer-tools-integrations
tags	["codex","claude-code","goal","prompt-engineering","agent-skills","verification"]
argument-hint	[vague-task-or-goal]
allowed-tools	Read, Bash(python *), Bash(py *)

Goal Meta Skill

把一个模糊任务，收敛成 Claude Code 或 Codex 可以持续执行、可以验证、知道何时停止和何时暂停的 /goal 指令。

In the commands below, <skill-dir> is this skill's base directory, announced when the skill loads. Substitute the literal path; it is not an environment variable. The script self-locates, so only the path to it must resolve.

Operating Mode

Run as a production-lite meta skill.

Shared defaults:

• The user wants a paste-ready /goal command for their platform, not a general prompt.
• The executable slash command stays /goal on both platforms. Do not output /目标 as the command unless the user's environment explicitly documents that alias.
• Goal objectives/conditions must be non-empty and at most 4,000 characters on both platforms. For longer instructions, draft a compact /goal that points at a local instruction file and put the long contract in that file content or a separate copy block.
• The first goal block should be the best recommended executable version, not a half-filled template. Users often copy the first draft directly.
• For Chinese users, output Chinese content and Chinese field names by default while keeping the command prefix /goal.
• For Chinese users, include both 推荐执行版（中文，可直接复制） and Goal Draft (English-compatible) unless the user asks for one language only.
• If the task is still vague but low-risk, choose the best conservative defaults and continue. Ask only when the answer changes cost, risk, ownership, or product direction.
• If the domain is unfamiliar or specialized, create a discovery-first goal that makes the agent inspect authoritative project/docs/runtime evidence before implementation instead of inventing domain rules.
• If the missing detail is low-risk, make an explicit assumption and continue.
• Do not start the work described by the goal unless the user explicitly asks. This skill creates the goal instruction.
• Prefer concrete verification commands and artifacts over vague confidence phrases.
• Prefer narrow write boundaries and explicit forbidden paths over broad permission.
• Treat Stop when and Pause if as part of the same completion/blocking contract.

Platform selection (details in references/platform-goal-facts.md):

1. If the user names Codex or Claude Code, render for that platform.
2. Otherwise infer from the host environment this skill is running in.
3. If still ambiguous, add one numbered choice such as 0. 平台：A Claude Code / B Codex to the existing 可选调整 block instead of opening a separate questionnaire round.

Codex-specific assumptions:

• /goal <text> sets a persistent thread objective; /goal views it; /goal pause, /goal resume, and /goal clear manage it. For inspect/pause/resume/clear requests, give the relevant management command instead of drafting a new objective.
• If /goal is missing from the slash command list, tell the user to enable features.goals in Codex config or run codex features enable goals.

Claude Code-specific assumptions:

• /goal <condition> sets a session completion condition and immediately starts a turn; an independent small model reads the transcript after each turn and auto-continues until the condition holds. /goal shows status; /goal clear (aliases stop, off, reset, none, cancel) removes it. There is no /goal pause or /goal resume — for a pause request, explain that and offer /goal clear plus re-setting later, or interrupting the session.
• Phrase the goal as a completion condition whose proof lands in the transcript: command exit codes, test/lint output, git status/git diff, file listings. The evaluator runs no tools, so screenshot-only or human-confirmation evidence must be rewritten into a transcript-visible check or moved into constraints.
• Always include a bounding clause such as 否则在 20 轮后停止并总结剩余问题 / or stop after 20 turns and summarize remaining issues.
• Keep the 暂停条件 / Pause if label but write its body as stop-and-report: Claude Code cannot pause, so high-risk situations must instruct Claude to stop, report, and wait for a human decision.
• Requires Claude Code v2.1.139+, an accepted workspace trust dialog, and hooks not disabled (disableAllHooks / allowManagedHooksOnly block it).

Workflow

1. Restate the task as an outcome, not an activity.
2. Determine the target platform using the platform selection rules above.
3. Classify the task using references/default-goal-strategy.md: familiar vs unknown domain, low vs high risk, new work vs existing project.
4. Choose best defaults for low-risk unknowns and write a one-sentence reason.
5. Identify missing information across the Goal contract:
   • success criteria
   • verification commands, artifacts, or evidence
   • constraints that must not change
   • allowed writes and forbidden paths
   • iteration policy
   • completion evidence
   • blocked stop conditions, human decisions, or budget caps
6. If the task is under-specified, prefer numbered multiple-choice adjustments with defaults. Use references/interview-checklist.md.
7. For Chinese-first users, produce the Chinese recommended execution goal first, then an English-compatible mirror that preserves the same meaning and keeps English field labels.
8. Render the contract for the target platform using the rendering rules in references/platform-goal-facts.md, then check the command against references/goal-command-playbook.md.
9. If the task is about an existing active goal, answer with the platform's management commands: Codex supports /goal, /goal pause, /goal resume, /goal clear; Claude Code supports /goal and /goal clear only.
10. Keep the executable /goal objective/condition within the 4,000 character limit. When the contract is longer, use a file-pointer pattern such as /goal Follow the task contract in .planning/<task>.md and stop only when its verification section is satisfied.
11. For file deliverables, run python "<skill-dir>/scripts/lint_goal_command.py" --platform <codex|claude> <file> or py -3 "<skill-dir>/scripts/lint_goal_command.py" --platform <codex|claude> <file> before calling the goal done. Add --require-chinese-companion when validating Chinese-first output.

Output Contract

When enough information is known, output the best recommended command first. Do not leave placeholders in real output.

Codex rendering (objective style):

/goal Create a first-version local MVP for the requested task, inspect project-provided commands before changing code, implement the core user-visible workflow, and keep unrelated systems unchanged.
Verification: run the smallest project-provided checks, start the local app or relevant runtime, complete the core workflow once, and capture logs/screenshots or command output as evidence.
Constraints: do not add accounts, paid services, production changes, destructive operations, or unrelated features unless requested.
Boundaries: write only inside the new project directory or the directly related existing project files.
Iteration policy: implement one focused workflow at a time, rerun checks after meaningful changes, inspect logs before retrying, and make at most 3 focused improvement rounds before reporting remaining risks.
Stop when: the core workflow is proven by runtime evidence and checks pass or missing checks are explicitly reported.
Pause if: credentials, payments, production data, destructive changes, legal/medical/financial decisions, copyrighted assets, or unclear ownership is required.

For Chinese-first Codex users, prefer this equivalent shape:

/goal 基于用户需求创建第一版本地 MVP，先读取项目已有命令和约束，实现核心用户可见流程，并避免改动无关系统。
验证：运行项目提供的最小相关检查，启动本地应用或对应运行环境，完整走通一次核心流程，并用日志、截图或命令输出作为证据。
约束：不加入账号、付费服务、生产变更、破坏性操作或无关功能，除非用户明确要求。
边界：只写入新项目目录，或只修改现有项目中与该功能直接相关的文件。
迭代策略：一次实现一个聚焦工作流，每次有意义改动后重跑检查，重试前先读日志，最多做 3 轮聚焦改进后报告剩余风险。
完成条件：核心流程有运行证据证明可用，检查通过或明确说明缺少配置。
暂停条件：需要凭证、付费、生产数据、破坏性操作、法律/医疗/金融判断、版权素材或所有权不清时暂停。

Claude Code rendering of the same request (condition style — proof lands in the transcript, bounded turns, stop-and-report instead of pause):

/goal 第一版本地 MVP 完成：核心用户可见流程可以完整走通一次，项目提供的最小相关检查命令退出码为 0 并把输出贴进对话，git status 显示只有新项目目录或直接相关文件被改动；否则在 20 轮后停止并总结剩余问题。
验证：运行项目提供的最小相关检查并展示命令退出码，启动本地应用或对应运行环境，完整走通一次核心流程，把检查输出、运行日志和 git status 结果贴进对话作为证据。
约束：不加入账号、付费服务、生产变更、破坏性操作或无关功能，除非用户明确要求。
边界：只写入新项目目录，或只修改现有项目中与该功能直接相关的文件。
迭代策略：一次实现一个聚焦工作流，每次有意义改动后重跑检查，重试前先读日志，最多做 3 轮聚焦改进后报告剩余风险。
完成条件：核心流程的检查输出和运行证据出现在对话记录中，检查通过或明确说明缺少配置。
暂停条件：需要凭证、付费、生产数据、破坏性操作、法律/医疗/金融判断、版权素材或所有权不清时，停止并报告，等待人工决定。

When the task is vague, output:

1. 推荐执行版（中文，可直接复制）: the best default /goal for the target platform.
2. 默认选择理由: one concise sentence.
3. 可选调整: numbered choices with recommended defaults and short option labels (include the platform choice only when the platform is ambiguous).
4. 你可以直接回复: an example such as 按默认 or 1B 2A 3C.
5. Goal Draft (English-compatible): a faithful English-compatible mirror with English field labels.

If the user writes in English, output only the English-compatible draft unless they ask for Chinese too.

If the user asks how to manage an existing goal, answer with the minimal platform-correct command:

• Codex: /goal to view, /goal pause to pause, /goal resume to continue, /goal clear to remove.
• Claude Code: /goal to view status, /goal clear to remove (aliases: stop, off, reset, none, cancel). There is no pause; to pause in spirit, clear now and re-set the same condition later, or interrupt the session.

Do not output long generic coaching unless the user asks for explanation.

Quality Bar

A strong goal:

• has one concrete outcome
• names exact checks or evidence
• protects unrelated files, user data, secrets, and default branches
• defines the write boundary
• tells the agent how to iterate after failures
• says when to stop because completion is proven
• says when to pause (Codex) or stop-and-report (Claude Code) because a human decision, credential, account state, budget, or repeated blocker is required
• fits inside the shared 4,000 character limit, or points to a file that contains the longer contract
• on Claude Code: reads as a completion condition, cites transcript-visible proof, and includes a turn/time bounding clause

Reject or revise a goal that:

• says only make it better, finish this, or fix bugs
• lacks verification
• lets the agent edit the whole machine or repo without reason
• asks for repeated retries without a new source of evidence
• has no pause/stop condition for external auth, secrets, payments, destructive actions, or ambiguous product decisions
• leaves placeholders such as [Outcome] in user-facing executable drafts
• treats vague words such as 高级, 有质感, or professional as verification instead of translating them into screenshots, runtime checks, review criteria, or iteration rules
• recommends /goal pause or /goal resume to a Claude Code user
• on Claude Code: relies on evidence the transcript evaluator cannot see, or has no bounding clause

Reference Files

• references/platform-goal-facts.md: single source of truth for Codex vs Claude Code /goal behavior, rendering rules, and platform selection.
• references/goal-command-playbook.md: the core /goal template, when to use it, examples, and anti-patterns.
• references/default-goal-strategy.md: lazy-user defaults, unknown-domain discovery, risk classification, and direct-copy output rules.
• references/interview-checklist.md: question bank for turning vague tasks into strong goals.
• scripts/lint_goal_command.py: lightweight checker for required /goal labels, unresolved placeholders, platform-specific rules (--platform codex|claude|both), the 4,000 character limit, and optional Chinese-first companion sections.