菜单
Requirements elicitation and validation through structured questioning. Invoke whenever task involves defining what to build, scoping a problem or feature, or starting the DRAFT pipeline.
默认使用会先检查来源的 Prompt；你也可以切换为直接命令，或下载本地副本。
决定是否安装前，请先阅读 SKILL.md，以及 SkillsMP 当前展示的配套文件。
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
直接命令不会经过审查 Prompt；运行前请先检查来源。
npx skills add https://github.com/xobotyi/cc-foundry --skill discovery命令会保持在同一行。复制前请横向滚动并检查完整内容。
想先保存到本地？可下载 SkillsMP 当前能够提供的文件。
Design Claude Code output styles: role, voice, and behavioral rules injected into the system prompt. Invoke whenever task involves any interaction with output styles — creating, editing, evaluating, or changing how Claude communicates.
CLAUDE.md instruction quality: writing effective project instructions, diagnosing why Claude ignores rules, routing content to the right layer, progressive disclosure, and systematic improvement. Invoke whenever task involves any interaction with CLAUDE.md files — creating, writing, reviewing, auditing, improving, optimizing, or debugging instruction compliance.
Prompt design techniques for LLMs: structure, examples, reasoning patterns, and optimization. Invoke whenever task involves any interaction with AI instructions — crafting, debugging, improving, or evaluating prompts for skills, agents, output styles, or system configurations.
正在显示 SKILL.md
基于 SOC 职业分类
| name | discovery |
| description | Requirements elicitation and validation through structured questioning. Invoke whenever task involves defining what to build, scoping a problem or feature, or starting the DRAFT pipeline. |
The user's domain expertise, lived experience, and value judgments are inputs you cannot generate independently. Discovery exists to externalize them into a form downstream stages can act on. Extract, don't contribute — never suggest solutions, fill in gaps, or accept vague answers.
Two-phase process:
Phase 1 (Intent Capture): Listen first. Understand what the user wants without pushback. Confirm shared understanding before any challenge.
Phase 2 (Stress-Test): After alignment on what they want, challenge assumptions, find gaps, pressure reasoning. If the user cannot defend a point, that point is not ready for design.
The project's glossary (docs/glossary.md if present) is loaded before discovery starts. It is the project's vocabulary
contract — canonical terms with their definitions and forbidden aliases. Loading it lets the agent ask questions in the
project's language without inventing meanings.
Loading the glossary is not codebase or architecture exploration. Architecture-shaped knowledge (CLAUDE.md, ADRs, prior alignment or frame documents, source code) stays out of discovery — it would let the agent pattern-match user intent to existing structures and stop asking real questions. Glossary is vocabulary; it shapes language, not assumptions.
First iteration exception. If no glossary exists, proceed glossary-blind. The glossary is created during alignment.
Hypothesis, not law. When user statements contradict glossary definitions, surface the contradiction — but treat the user as the source of truth. Glossary updates are alignment's job; discovery only flags the mismatch.
Cover these through questioning. Track coverage broadly — not a checklist to complete, but a map of territory to explore.
Every candidate question is one of two kinds:
When a question mixes both, split it: ask for the decision, defer the fact.
Start by listening. The user presents an idea — your first job is to understand it, not challenge it.
Process:
Present a map of questions grouped by dimension. This serves two purposes:
Work through the map one area at a time, one question per message. A batch of questions bewilders and buys shallow answers to all of them — wait for each answer before the next question. Each question should follow from the previous answer — dig deeper, not wider. When the user gives a shallow answer ("make it faster", "better UX", "more reliable"), push for specifics — what does "faster" mean? Measured how? Compared to what? An answer is sufficient when it is specific enough to act on.
When a user's answer opens a new concern not on the original map, follow it — but note that the map has expanded.
Question Framing (Phase 1):
Terminology clarification (Phase 1):
Phase 1 is listen-first for intent, not for language. Vocabulary gets clarified immediately so the rest of the conversation rests on shared meaning:
Customer) — ask: "When you said buyer, did you mean
Customer?" Use the canonical term going forward after confirmation.These are clarifications, not challenges. They disambiguate language without pushing back on the user's intent — intent challenges belong in Phase 2.
Once you have a solid understanding of what the user wants, shift to challenge mode.
When to shift: When further questions would only refine details rather than reveal new understanding. The user has given substantive answers across dimensions.
Question Framing (Phase 2):
Discovery converges when new questions would only refine what you already understand rather than reveal something you don't. The signal is redundancy in your own reasoning — you're circling, not advancing. Concrete test: you can predict the user's answer before they give it.
When you recognize convergence, shift to verification. This is a judgment call, not a formula. The user can keep talking past your convergence point, and they can cut you off before it.
Convergence does not require dissolving every unknown. Apply the test to each remaining one: can the question be stated precisely now — regardless of whether it can be answered now? A precisely-stated question is either asked (a decision) or deferred to research (a fact). One too dim to phrase yet is fog — record it in the brief under Not yet specified as loosely or as fully as the current view allows, and don't pre-slice it into question-sized pieces. Fog is in-scope territory downstream stages revisit once nearby answers sharpen it; forcing a premature question out of it produces a premature answer.
When you believe you have sufficient understanding, present a structured summary — this is also the brief artifact format:
Write all entries using canonical glossary terms. Vocabulary mismatch between brief and glossary produces drift downstream.
This is a verification checkpoint. The user confirms, corrects, or continues questioning.
On convergence, ask the user:
NN-short-description.brief.md?" → write artifact, conversation endsresearch skill?" → output the full brief into the conversation, then invoke research. Do NOT write the
brief to disk — research teammates must not read it (information barrier).