plannotator-setup-goal
关于
This Claude Skill transforms a user's idea into an executable `/goal` package by conducting a structured interview and building a reviewed fact sheet. It then explores the codebase to create a detailed execution plan, saving all artifacts to a dedicated `goals/<slug>/` directory. Use it to systematically define and scope new development objectives from initial concepts.
快速安装
Claude Code
推荐npx skills add backnotprop/plannotator -a claude-code/plugin add https://github.com/backnotprop/plannotatorgit clone https://github.com/backnotprop/plannotator.git ~/.claude/skills/plannotator-setup-goal在 Claude Code 中复制并粘贴此命令以安装该技能
技能文档
Setup Goal
Turn an idea into a goal package at goals/<slug>/ through structured discovery, user interview, and codebase exploration.
Phases
1. Rearticulate
State back what the user wants in your own words. If the conversation already has rich context, summarize it. If the goal is bare or vague, do minimal shallow exploration of the codebase to ground your understanding. Keep it to 2-3 sentences. Wait for the user to confirm or correct before continuing.
Create the goal directory once the slug is clear:
mkdir -p goals/<slug>
Use goals/<slug>/ for both working JSON files and final docs. The JSON files are provenance and iteration state; the markdown files are the human-readable authoritative goal package.
Browser session patience rule: Plannotator goal setup is a user-driven browser session. After launching an interview or facts command, be absolutely patient and keep waiting on the user until they submit, dismiss, or explicitly ask you to stop. Do not close, kill, restart, refresh, or open a second copy just because the UI is idle or the user is taking time. Never close and reopen the session as a way to update state; if a rerun is needed after the prior session ends, update the working JSON file and launch a new command from that file.
Optional: grill first (deep, one-at-a-time interview). Before building the compact interview bundle, suggest a grilling pass whenever the goal is vague or carries many interdependent decisions — and run one whenever the user asks for it ("grill me first"). This is opt-in: for a clear, well-scoped goal, skip it and go straight to the bundle, so grilling never fights the bundle's "fewer, higher-leverage questions" philosophy. When you grill, run the protocol below verbatim, then fold the resolved decisions forward into a higher-quality interview bundle (Phase 2) — or, if grilling fully resolves scope, straight into the fact sheet (Phase 3).
<!-- Grilling protocol below adapted verbatim from the /grill-me skill by Matt Pocock (MIT-licensed): https://github.com/mattpocock/skills/blob/main/skills/productivity/grill-me/SKILL.md -->Interview me relentlessly about every aspect of this plan until we reach a shared understanding. Walk down each branch of the design tree, resolving dependencies between decisions one-by-one. For each question, provide your recommended answer.
Ask the questions one at a time.
If a question can be answered by exploring the codebase, explore the codebase instead.
2. Interview Bundle
Build a compact bundle of questions that can derive every "fact" this goal should produce. Package the questions together so the user can answer them quickly in the Plannotator goal setup UI. For each question, include your recommended answer and use options when they make answering faster.
Do not ask obvious confirmation questions. If the answer can be inferred from the user's request, from the conversation, or from shallow codebase exploration, infer it and move on. If an obvious area has meaningful nuance, present the inferred answer as a recommendation with options or a custom "add/correct this" path rather than asking the user to restate the obvious.
Question areas that usually matter:
- What the feature/change is
- Who it's for
- What problem it solves
- What behavior changes
- What success looks like
- What's in and out of scope (the most important area to determine facts)
- What edge cases to consider
- What constraints or precedent apply
If a question can be answered by exploring the codebase, explore the codebase instead of asking. Only include questions where the user's judgment is actually needed. Prefer fewer, higher-leverage questions over exhaustive obvious ones.
Write the interview bundle before showing it to the user:
goals/<slug>/interview.json
{
"stage": "interview",
"title": "Short human-readable title",
"goalSlug": "<slug>",
"questions": [
{
"id": "scope",
"prompt": "What should be in scope?",
"description": "Optional clarification.",
"answerMode": "multi-custom",
"recommendedAnswer": "Your recommended answer.",
"recommendedOptionIds": ["ui", "server"],
"options": [
{ "id": "ui", "label": "UI" },
{ "id": "server", "label": "Server" }
],
"required": true
}
]
}
Supported answerMode values: text, single, multi, custom, single-custom, multi-custom.
Run this as a monitored foreground process and wait patiently for the browser session to finish. The command may appear idle while the user is reading, editing, or asking questions; leave it running:
plannotator setup-goal interview goals/<slug>/interview.json --json
The command returns JSON on stdout with the submitted answers. Write that exact result to goals/<slug>/interview-result.json before continuing. A convenient pattern is:
plannotator setup-goal interview goals/<slug>/interview.json --json | tee goals/<slug>/interview-result.json
If the user revises after the session finishes, update interview.json and rerun the command instead of reconstructing the whole bundle from memory. If the session is dismissed, stop and tell the user the goal setup was closed.
Before moving to facts, read every answer and note carefully:
- If the user wrote questions, uncertainty, "not sure", "needs context", or similar concerns in an answer or note, stop and address those questions in chat. Do not proceed to facts until the user has enough context or you have rerun a revised interview bundle.
- If the user skipped a question with a note, treat the note as intentional feedback, not as an empty answer. Answer the note, refine the question, or make a documented assumption before proceeding.
- If the user skipped a question without a note, proceed only if the missing answer is non-blocking; otherwise ask the smallest possible follow-up in chat.
3. Fact Sheet
A fact is a simple description of each outcome of a goal. It should be easily testable and verifiable. A fact may describe the function of a specific feature or aspect of a system. A fact may determine specific UI and UX. Again, a fact is literally anything that can be tested and verified in automated or manual testing. Keep fact language simple. In a way, a fact sheet is a design spec, but less verbose & using language the human user can easily visualize & rationalize.
Prepare a facts review bundle from goals/<slug>/interview-result.json. Each fact should include whether automated verification is recommended and preselected.
Write the facts review bundle before showing it to the user. If revising after a prior facts pass, start from facts-review.json and facts-result.json, include previously accepted facts with "accepted": true, and preserve their state.
goals/<slug>/facts-review.json
{
"stage": "facts",
"title": "Short human-readable title",
"goalSlug": "<slug>",
"facts": [
{
"id": "fact-1",
"text": "The accepted fact text.",
"accepted": false,
"removed": false,
"recommendedAutomatedVerification": true,
"automatedVerification": true
}
]
}
Run this as a monitored foreground process and wait patiently for the browser session to finish. The command may appear idle while the user is reviewing, editing, or asking questions; leave it running:
plannotator setup-goal facts goals/<slug>/facts-review.json --json
The command returns JSON on stdout with accepted/edited/removed facts plus automated verification selections. Write that exact result to goals/<slug>/facts-result.json. A convenient pattern is:
plannotator setup-goal facts goals/<slug>/facts-review.json --json | tee goals/<slug>/facts-result.json
Write goals/<slug>/facts.md as a flat readable list of accepted facts. Each fact is one line; add a minimal note only when the fact cannot be stated clearly on its own. Also write goals/<slug>/facts.meta.json preserving each accepted fact's id, final text, comment, recommendedAutomatedVerification, and automatedVerification value.
If the user edits or removes facts in the UI, apply that result directly. If the session is dismissed, stop and tell the user the facts review was closed.
4. Plan
Explore the codebase. Discover and validate implementation paths toward each accepted fact. Treat facts with automatedVerification: true as requiring concrete automated checks unless you document a blocker. Trace through code, identify files and systems involved, surface risks and unknowns. Refine until you have a confident order of operations.
Write goals/<slug>/plan.md:
- Solution approach (brief)
- Ordered steps with the files/systems each touches
- Verification for each step (concrete commands or checks)
- Risks or open questions worth flagging
Gate the plan with Plannotator:
plannotator annotate goals/<slug>/plan.md --gate
If denied, revise from feedback and re-gate until approved.
5. Goal Output
Write goals/<slug>/goal.md:
- The articulated goal (1-3 sentences)
- Reference to
facts.mdas the shared understanding - Reference to
plan.mdas the execution plan - Done condition
Tell the user:
Done! Launch a goal with `/goal goals/<slug>/goal.md`
GitHub 仓库
相关推荐技能
content-collections
元Content Collections 是一个 TypeScript 优先的构建工具,可将本地 Markdown/MDX 文件转换为类型安全的数据集合。它专为构建博客、文档站和内容密集型 Vite+React 应用而设计,提供基于 Zod 的自动模式验证。该工具涵盖从 Vite 插件配置、MDX 编译到生产环境部署的完整工作流。
polymarket
元这个Claude Skill为开发者提供完整的Polymarket预测市场开发支持,涵盖API调用、交易执行和市场数据分析。关键特性包括实时WebSocket数据流,可监控实时交易、订单和市场动态。开发者可用它构建预测市场应用、实施交易策略并集成实时市场预测功能。
creating-opencode-plugins
元该Skill帮助开发者创建OpenCode插件,用于接入命令、文件、LSP等25+种事件。它提供了插件结构、事件API规范和JavaScript/TypeScript实现模式,适合需要拦截操作、扩展功能或自定义事件处理的场景。开发者可通过它快速构建响应式模块来增强OpenCode AI助手的能力。
sglang
元SGLang是一个专为LLM设计的高性能推理框架,特别适用于需要结构化输出的场景。它通过RadixAttention前缀缓存技术,在处理JSON、正则表达式、工具调用等具有重复前缀的复杂工作流时,能实现极速生成。如果你正在构建智能体或多轮对话系统,并追求远超vLLM的推理性能,SGLang是理想选择。
