goal

Goal

This skill creates the /goal instruction before execution starts. It does not perform the requested work unless the user explicitly asks to continue after the goal is written.

Operating Rules

• Keep the command prefix exactly /goal, even for Chinese users. The body may be Chinese.
• Put the best copy-ready goal first. Do not start with placeholders or a worksheet.
• Choose conservative defaults for low-risk ambiguity and state the assumption briefly.
• Ask only when the answer changes risk, cost, ownership, product direction, or permission.
• For specialized or unfamiliar domains, make the goal discovery-first instead of inventing domain rules.
• Treat verification, write boundaries, stop conditions, and pause conditions as required parts of the contract.
• Prefer commands, logs, screenshots, files, URLs, builds, tests, or other evidence over vague confidence.

Goal Contract

Every finished goal should contain these elements:

/goal [one concrete outcome].
Verification: [commands, runtime checks, artifacts, screenshots, logs, files, URLs, or other evidence].
Constraints: [behavior, data, APIs, style, safety, branch, or policy rules that must not change].
Boundaries: [allowed write paths and explicitly forbidden paths or systems].
Iteration policy: [small steps, rerun checks, inspect evidence before retrying, retry cap when useful].
Stop when: [specific evidence proves completion].
Pause if: [credentials, production data, destructive actions, external approval, legal/medical/financial judgment, budget, ownership, or repeated blockers are required].

For Chinese users, use Chinese labels:

/goal [一个具体目标结果]。
验证：[命令、运行检查、产物、截图、日志、文件、链接或其他证据]。
约束：[不能改变的行为、数据、接口、风格、安全、分支或策略规则]。
边界：[允许写入的位置，以及明确禁止触碰的路径或系统]。
迭代策略：[小步推进，重跑检查，重试前先看证据，必要时限制轮数]。
完成条件：[哪些具体证据证明可以停止]。
暂停条件：[需要凭证、生产数据、破坏性操作、外部审批、法律/医疗/金融判断、预算、所有权或重复阻塞时暂停]。

Drafting Flow

1. Convert the request into an outcome, not an activity.
2. Classify risk: local prototype or docs is low risk; existing repo changes are medium risk; production data, credentials, payments, compliance, legal, medical, financial, destructive operations, live deployment, or copyrighted assets are high risk.
3. If the project exists, require the agent to inspect project scripts, docs, tests, and conventions before inventing commands.
4. Write verification that produces evidence. If exact commands are unknown, make command discovery part of verification.
5. Separate constraints from boundaries: constraints preserve behavior and policy; boundaries define where the agent may write.
6. Bound iteration: one focused change at a time, rerun checks, inspect logs before changing strategy, and avoid endless retries.
7. Define stop and pause conditions so the agent knows when to finish and when to ask.

Output Shapes

For a vague Chinese request, output:

1. 推荐执行版（中文，可直接复制）
2. one short 默认选择理由
3. 可选调整 with two to four numbered choices and default options
4. 你可以直接回复：按默认，或回复类似 1B 2A

For an English request, output one English-compatible goal unless the user asks for Chinese too.

For a high-risk or unfamiliar task, output a discovery-first goal that verifies available docs, samples, scripts, policies, or official references before implementation and pauses before the risky action.

Quality Checks

Reject or revise a draft when it:

• does not start with /goal
• lacks verification, constraints, boundaries, iteration policy, stop conditions, or pause conditions
• contains placeholders such as [Outcome], TODO, TBD, 待定, or 待补充
• uses weak verification like make sure it works or 确认可用 without evidence
• allows broad writes like edit anything, 随便改, or 修改任意文件
• says to retry forever without a new source of evidence
• treats vague taste words such as 高级, 专业, or polished as completion proof instead of translating them into screenshot or review evidence

When saving a goal to a file, run:

python3 "${CLAUDE_SKILL_DIR}/scripts/check_goal_contract.py" path/to/goal.txt

The checker lives at scripts/check_goal_contract.py.