optimize-agent-docs（优化 Agent 文档）

Agent Knowledge Optimizer

Transform accumulated documentation into a retrieval-optimized knowledge system.

Core Principle

File organization is a human concern. Agents don't browse—they search and load. Optimize for:

• Discovery: What knowledge exists?
• Relevance: Is it needed for this task?
• Efficiency: What's the minimum to load?

Workflow

Phase 1: Knowledge Extraction

Inventory all agent documentation:

# Find all agent doc sources
find . -maxdepth 2 -name "*.md" -path "*/.claude/*" -o \
       -name "*.md" -path "*/.codex/*" -o \
       -name "*.md" -path "*/.cursor/*" -o \
       -name "CLAUDE.md" -o -name "AGENTS.md" -o -name "INSTRUCTIONS.md"

For each file, extract:

• Discrete facts (single pieces of actionable information)
• Instructions (procedures, rules, constraints)
• Context triggers (when is this knowledge needed?)

Phase 2: Chunk Analysis

Break content into retrieval units—the smallest self-contained piece of information that makes sense alone.

Good chunk:

## Adding API Endpoints
1. Create handler in src/handlers/
2. Register route in src/routes.rs
3. Add OpenAPI spec to docs/api.yaml

Bad chunk (too coupled):

See the API section for endpoint patterns, but first read the auth docs,
which reference the middleware guide...

Score each chunk:

• Self-contained? Can agent act on this without loading more?
• Task-specific? Clear when this is needed?
• Information-dense? High signal per token?

Phase 3: Build Knowledge Manifest

Generate .claude/KNOWLEDGE.md—a lightweight index the agent reads first:

# Knowledge Manifest

## Task → Knowledge Map

| When working on... | Load | Key terms |
|-------------------|------|-----------|
| API endpoints | references/api.md | route, handler, endpoint |
| Authentication | references/auth.md | token, session, login |
| Database changes | references/schema.md | migration, model, query |
| Testing | references/testing.md | spec, fixture, mock |
| Deployment | references/deploy.md | release, staging, prod |

## Quick Reference

### Build Commands
- `npm run dev` — Start dev server (port 3000)
- `npm test` — Run test suite
- `npm run build` — Production build

### Key Paths
- Handlers: `src/handlers/`
- Routes: `src/routes.ts`
- Tests: `tests/`

### Critical Rules
- Never commit .env files
- All PRs require tests
- Use conventional commits

The manifest contains:

1. Task→Knowledge map: What to load for what context
2. Quick reference: High-frequency facts (no file loading needed)
3. Critical rules: Must-know constraints (always relevant)

Phase 4: Compile Optimized Artifacts

Transform verbose source docs into dense, agent-optimized versions.

Compression techniques:

Output structure:

.claude/
├── CLAUDE.md              # Human-readable, can stay verbose
├── KNOWLEDGE.md           # Agent manifest (generated)
└── compiled/              # Agent-optimized versions (generated)
    ├── api.md             # Dense API reference
    ├── patterns.md        # Code patterns as templates
    └── rules.md           # All constraints in one place

Phase 5: Generate Retrieval Hints

Add grep-friendly markers throughout compiled docs:

<!-- @task:new-endpoint @load:api,routes -->
## Adding Endpoints

<!-- @task:fix-auth @load:auth,middleware -->
## Authentication Flow

<!-- @task:write-test @load:testing -->
## Test Patterns

These markers enable:

# Find relevant sections for a task
grep -l "@task:new-endpoint" .claude/compiled/*.md

Phase 6: Validation

Test the optimized system:

1. Coverage check: Every fact from source exists in compiled output
2. Retrieval test: Can common tasks be served with minimal loading?
3. Density check: Compiled versions smaller than sources?

# Compare sizes
wc -l .claude/references/*.md    # Source
wc -l .claude/compiled/*.md       # Compiled (should be smaller)

Manifest Format

The KNOWLEDGE.md manifest follows this structure:

# Knowledge Manifest
<!-- Auto-generated. Source: .claude/references/, CLAUDE.md -->

## Task Context Map
<!-- What to load based on current work -->

| Context | Load | Search |
|---------|------|--------|
| [task description] | [file path] | [grep terms] |

## Always-Loaded Facts
<!-- High-frequency, never needs file lookup -->

### Commands
[Most-used commands as a table]

### Paths
[Key directories and their purposes]

### Rules
[Critical constraints that always apply]

## Chunk Index
<!-- What exists and where -->

| Topic | Location | Lines | Summary |
|-------|----------|-------|---------|
| [topic] | [file:line-range] | [count] | [one-line summary] |

Information Density Principles

Convert Prose to Structure

Before:

"The authentication system uses JWT tokens stored in httpOnly cookies. When a user logs in, the server validates credentials against the database, generates a token with a 24-hour expiry, and sets it as a cookie..."

After:

## Auth Flow
- Method: JWT in httpOnly cookie
- Expiry: 24h
- Flow: credentials → DB validate → token → cookie

Eliminate Redundancy

If the same information appears in multiple places, create one canonical source and reference it:

## Token Handling
See: [Auth Flow](#auth-flow) — tokens section

Prefer Tables Over Lists

Before:

- The API endpoint for users is /api/users
- The API endpoint for posts is /api/posts
- The API endpoint for comments is /api/comments

After:

| Resource | Endpoint |
|----------|----------|
| Users | /api/users |
| Posts | /api/posts |
| Comments | /api/comments |

Use Patterns Over Examples

Before:

To create a user handler:

export async function createUser(req, res) { const { name, email } = req.body; const user = await db.users.create({ name, email }); res.json(user); }

After:

Handler pattern: export async function {action}{Resource}(req, res) Body: Extract params → DB operation → Return result

## Output Checklist

After optimization, verify:

- `KNOWLEDGE.md` exists and is under 100 lines
- Task→knowledge mappings cover common workflows
- Quick reference has most-used facts
- Compiled docs are denser than sources
- No orphaned knowledge (everything indexed)
- Retrieval hints enable grep-based discovery
- Original source docs untouched (human reference)