Goal Prompt Builder

/goal Prompt Builder

This skill turns a fuzzy task description ("我想让 Codex 帮我重构鉴权") into a complete, audit-friendly /goal command that's ready to paste into Codex CLI 0.128+.

Why this skill exists

Codex 0.128 added /goal as a persistent objective with a runtime-injected audit prompt (continuation.md). The audit prompt forces the model to build a "prompt-to-artifact checklist" — but only if the user's goal text can be mapped to one. Vague goals produce vague checklists, which produce false completions. This skill exists to make sure every /goal you generate has the structure that lets the audit mechanism actually work.

The golden template (5 sections, in this order)

Every /goal Claude generates with this skill follows this structure exactly:

/goal <objective>.

[Optional: First action: read X, Y, Z and report counts. Wait for ack.]

Scope: <files / subsystem / feature area>.

Constraints:
  - <what not to change>
  - <compatibility / permission boundaries>
  - <project-specific rules from AGENTS.md / CLAUDE.md>

Done when:
  1. <verifiable artifact 1 — cite file or command>
  2. <verifiable artifact 2>
  ...

Stop if:
  - <mechanically detectable condition 1>
  - <mechanically detectable condition 2>
  ...

Use a token budget of <N> tokens for this goal.

Why this order: matches continuation.md's expected reading flow — objective first, then scope to bound the search, then constraints to prune options, then acceptance to define success, then stop-if as runtime guards.

Workflow

When this skill triggers, walk the user through these 6 steps. Step 0 (interaction mode) and Step 1 (project detection) happen automatically — Step 0 needs one user choice, Step 1 needs zero if filesystem is accessible. Don't skip steps unless the user explicitly says "I'll fill it in myself, just give me the template".

Step 0: Pick interaction mode

This skill supports three interaction modes. Ask once at the start:

你希望用哪种方式生成 /goal？

• A. 询问式 — 我一段一段问你（最稳，适合第一次写 /goal）
• B. 全描述式 — 你一句话描述需求，我拆解后只问你确认不确定的地方（最快，适合熟手）
• C. 混合式（默认） — 先选场景模板，再问 3-5 个关键问题（推荐）

Once chosen, follow the matching flow:

• A. 询问式 → Step 1 → Step 2 → Step 3a → 3b → 3c → 3d → 3e → Step 4 → Step 5 (each input gathered separately)
• B. 全描述式 → Step 1 → ask "用一段话描述你想做什么 / scope / 验收 / 不希望发生什么" → parse into 5 sections → Step 4 → ask user to confirm only the gaps → Step 5
• C. 混合式 → Step 1 → Step 2 → Step 3 (batched: ask all missing fields at once) → Step 4 → Step 5

Mode B is most powerful when the user has thought about the task. Mode A is safest when they haven't. Mode C is the default sweet spot.

If the user doesn't answer this question explicitly, default to mode C and proceed.

Step 1: Detect (don't ask) the project type

Auto-detection comes first. Only fall back to asking if detection fails.

Run this detection sequence:

1. Check the conversation context. Has the user already provided a repo URL, file path, or code snippet? Read those for hints first.

2. Probe the filesystem (if you have file tools and the user is in a project directory):

   • package.json exists → Node / TypeScript
   • pyproject.toml or requirements.txt or setup.py → Python
   • *.xcodeproj/ or Package.swift → Swift / iOS
   • Cargo.toml → Rust
   • go.mod → Go
   • astro.config.* / next.config.* / _config.yml / mkdocs.yml → Static / docs project

3. Fetch the repo if the user gave a URL (only if web tools available):

   • GitHub URL → fetch the README + try to identify config files
   • Read CLAUDE.md and AGENTS.md if they exist (these are gold for Constraints)

4. Fall back to asking only if all auto-detection failed:

   我没法自动判断项目类型——这是 Node / Python / Swift / Go / Rust / 静态 / 其他？

When detection succeeds, announce what you found in one sentence so the user can correct you:

检测到这是一个 Swift / iOS 项目（找到 lingolearn.xcodeproj + CLAUDE.md）。 我会按 Swift 项目默认约束适配。如果不对，告诉我。

Then load the matching reference from references/project-types.md. Also load any CLAUDE.md / AGENTS.md found — these contain project-specific rules that override defaults.

Step 2: Pick a scenario template

Ask: 这个 goal 属于哪种类型？

Each option maps to a different default skeleton — see references/scenarios.md for the full templates.

Step 3: Gather the 5 inputs (in this order)

Ask only what's still missing. Don't ask all 5 at once — ask incrementally so the user can think.

3a. Objective (一句话)

• One sentence describing what changes by the end.
• If the user gives a verb-less noun phrase ("Cohere rerank support"), turn it into a verb phrase ("Add Cohere rerank support to retrieval pipeline").
• Reject vague verbs like "improve", "optimize", "clean up" — ask for the concrete change.

3b. Scope (改什么、不改什么)

• Which files / directories / subsystems are in play
• For brownfield projects: probe whether v1.x beta files or sensitive modules exist that should be off-limits

3c. Constraints (硬约束)

• Pull from AGENTS.md / CLAUDE.md if available
• Add project-type defaults from the loaded reference (e.g., for Swift: "do not modify project.pbxproj")
• Ask if there's a "MUST NOT modify" list

3d. Done when (验收清单)

• This is the most important section. Push back hard if items are vague.
• Each item must cite a file, command, test name, or measurable artifact.
• Replace "测试通过" → "<exact command> exits 0; paste summary"
• Replace "做完" → enumerate the deliverables
• Aim for 5-8 items; fewer than 3 is a red flag

3e. Stop if (停止条件)

• Each must be mechanically detectable
• Include project-type defaults (e.g., for Node: "needs npm install for new dep")
• Include a regression guard: "existing tests start failing — do not fix by editing tests"

Step 4: Predict audit-friendliness

Before showing the final command, internally score it (don't show the math, just the verdict):

• Acceptance count: 0 = bad, 1-2 = warn, 3-5 = good, 6-8 = excellent
• Vague verbs detected: "improve", "optimize", "全部", "彻底", "all", "everything" → flag
• Stop-if specificity: "if unclear" = bad, "if file X appears in git diff" = good
• Token budget present: missing = warn
• Mechanical verifiability: every Done-when item has a cite-able artifact = good

If score is below ~70%, don't ship the command yet. Instead, surface the weak spots and ask the user to refine. Be specific:

⚠ 我发现三处可以加强的地方：

1. Done when 第 2 条"测试覆盖" → 改成"测试名称 + 退出码"会更准
2. Stop if 缺少"现有测试 regression"兜底
3. Token 预算未指定，建议 80K（基于 scope 大小）