AgentSkillsCN

clarify

审阅现有的 CLAUDE.md 文件,通过有针对性的问题识别其中的不足之处并加以完善;随后提出更广泛的项目相关问题,将学习心得记录于 MEMORY.md 中。可在 /cmd-generator 之后使用,或在 CLAUDE.md 看似不够完整时随时启用。触发词包括:“clarify”、“确认”、“深掘”、“CLAUDE.md 确认”、“了解更多”、“精炼 CLAUDE.md”。

SKILL.md
--- frontmatter
name: clarify
description: >
  Review an existing CLAUDE.md, identify gaps, and refine it through targeted questions.
  Then ask broader project questions to capture learnings in MEMORY.md.
  Use after /cmd-generator or whenever the CLAUDE.md feels incomplete.
  Triggers: "clarify", "確認", "深掘り", "CLAUDE.md確認", "もっと詳しく", "refine CLAUDE.md".

Clarify Skill

Refine a project's CLAUDE.md by identifying gaps, then capture broader project knowledge in MEMORY.md.

Workflow

Phase 1: Refine CLAUDE.md

1. Locate and Read CLAUDE.md

Search for the project's CLAUDE.md in this order:

  1. ./CLAUDE.md (current working directory)
  2. Nearest ancestor directory's CLAUDE.md
  3. If none found, tell the user to run /cmd-generator first and stop.

Read the file fully before proceeding.

2. Gap Analysis

Analyze the CLAUDE.md for weak areas. Check each section:

SectionLook for
CommandsMissing key commands (dev, build, test, lint, deploy), undocumented flags/options
ArchitectureImportant directories not explained, missing layers (API, DB, middleware)
ConventionsMissing naming patterns, import order, error handling, logging conventions
Do NOTCommon pitfalls not yet listed, footguns specific to the stack
TestingMissing test patterns, fixture setup, mocking approaches, coverage expectations
Key DecisionsUndocumented architectural choices that affect daily coding

Also scan the actual codebase (package.json, directory structure, config files) to find things the CLAUDE.md should mention but doesn't.

3. Ask Targeted Questions (Round 1)

Use AskUserQuestion with up to 4 questions about the weakest areas found in step 2. Focus on gaps that would most improve Claude's effectiveness in this project.

4. Update CLAUDE.md and Re-analyze

Apply the answers to the appropriate sections of CLAUDE.md. Show the user a summary of what was added or changed.

Then re-evaluate: are there still significant gaps? If so, proceed to Round 2. If CLAUDE.md is now solid, skip directly to Phase 2.

5. Ask Follow-up Questions (Round 2, if needed)

Use AskUserQuestion with up to 4 more questions, now informed by the Round 1 answers. Good candidates for Round 2:

  • Details that Round 1 answers revealed (e.g., user mentioned a test framework → ask about fixture conventions)
  • Sections still weak after Round 1
  • Cross-cutting concerns (error handling, logging, env config) not yet captured

Update CLAUDE.md with the answers and show what changed.

Phase 2: Broader Project Q&A

6. Ask Broader Questions (Round 3)

Ask up to 4 questions about topics NOT covered by CLAUDE.md:

  • Workflow preferences (branching strategy, PR process, CI/CD pipeline)
  • Team conventions (review process, deployment cadence, release tagging)
  • Known pain points or tech debt areas
  • Personal preferences (response language, verbosity, commit message style)

Skip any topic already well-documented in CLAUDE.md or MEMORY.md.

7. Classify and Write Answers

Classify each answer and write to the appropriate file:

ClassificationDestination
Project-specific conventions, architecture, commandsProject CLAUDE.md
Project-specific workflow, gotchas, team conventionsProject MEMORY.md
Personal preferences, general tools/habits~/.claude/MEMORY.md

Before creating any MEMORY.md file, ask the user for approval.

8. Show Summary

Present what was updated:

code
## Changes Made
- CLAUDE.md: added X, updated Y
- MEMORY.md (project): created with Z
- ~/.claude/MEMORY.md: appended W

Important Rules

  • Always read CLAUDE.md first — never ask about what's already documented
  • Use AskUserQuestion tool for all questions (not free-text), with concrete options when possible
  • Maximum 3 rounds of questions (4 per round, 12 total max) — Round 2 is conditional on remaining gaps
  • Always show what changed after each update
  • Do NOT create MEMORY.md files without explicit user approval
  • Do NOT ask generic questions — every question must be grounded in an observed gap
  • Scan the codebase to validate and supplement CLAUDE.md content (don't rely solely on the file)
  • Match the user's language (Japanese or English) based on the existing CLAUDE.md