discovery-first

krwhynot 的头像作者 krwhynot

使用预计算的发现文件于.claude/state/,用于代码库清单问题。触发于组件、钩子、模式、表单、类型、清单、存在什么、列出所有、有多少。比源扫描节省90%以上令牌。

📖 中文导读:使用预计算的发现文件于.claude/state/,用于代码库清单问题。触发于组件、钩子、模式、表单、类型、清单、存在什么、列出所有、有多少。比源扫描节省90%以上令牌。 以下正文为该技能的原始 SKILL.md 内容(英文),上方安装命令可直接使用。

name: discovery-first description: Use pre-computed discovery files at .claude/state/ for codebase inventory questions. Triggers on component, hook, schema, form, type, inventory, what exists, list all, how many. Saves 90%+ tokens vs source scanning.

Discovery-First Skill

Purpose

ALWAYS read discovery files before scanning source code for inventory questions.

When This Triggers

  • "What components exist in contacts?"
  • "List all hooks in the codebase"
  • "What schemas are used for opportunities?"
  • "How many forms exist?"
  • "Which types are defined?"
  • "What calls useContactData?"

Discovery Files Location

Question Type Read This First
Components .claude/state/component-inventory/{feature}.json
Hooks .claude/state/hooks-inventory.json
Schemas .claude/state/schemas-inventory/{feature}.json
Forms .claude/state/forms-inventory.json
Types .claude/state/types-inventory/{feature}.json
Call Graph .claude/state/call-graph-inventory/{feature}.json

Chunk Discovery Pattern

  1. Read manifest.json in the inventory directory first
  2. Find the chunk that contains the feature you need
  3. Read only that chunk (e.g., contacts.json)

Why This Matters

Method Token Cost Speed
Discovery JSON ~30KB < 2 sec
Source scanning ~500KB 5-15 sec
Savings 94% 5x faster

Required Behavior

  1. DO: Read .claude/state/ JSON files first
  2. DO: Use manifest.json to find the right chunk
  3. DON'T: Grep/Glob src/atomic-crm for inventory questions
  4. DON'T: Use Task/Explore agent for basic component/hook listing
  5. DON'T: Read individual source files when inventory answers the question

Example: Correct Approach

Question: "What hooks does ContactList use?"

Wrong approach:

Grep("use.*=", "src/atomic-crm/contacts/ContactList.tsx")
Read("src/atomic-crm/contacts/ContactList.tsx")

Correct approach:

Read(".claude/state/component-inventory/contacts.json")
→ Find ContactList entry → Read hooks array

Freshness

Discovery files are auto-regenerated on pre-commit via .husky/pre-commit. If data seems stale, run: just discover