documentation-inversion（文档倒置）

Documentation Inversion

Traditional documentation drifts out of sync because it lives separately from the code. Documentation inversion places guidance directly in the codebase, structured for agent consumption, so that AI can explore the source and find current information on demand.

Core Principle

Structure documentation for programmatic navigation -- hierarchical, cross-referenced, with explicit entry points -- so that AI agents can find what they need without human guidance.

Traditional documentation assumes a human reader who will browse, search, and interpret context. Agent-first documentation assumes a machine reader that needs: explicit entry points, structured hierarchies, cross-references it can follow programmatically, and guidance on what to explore next.

The Problem: Traditional Documentation Rots

Traditional Flow

Developer ships a feature → Writes a wiki article explaining it
                                      ↓
                               Weeks or months elapse
                                      ↓
                               Another developer refactors the feature
                                      ↓
                               The wiki article is never revised
                                      ↓
                               A third developer (or an agent) follows the stale article
                                      ↓
                               Incorrect assumptions, wasted effort, subtle bugs

Why it rots:

• Documentation is a second-class artifact -- the incentive is to ship code, not update docs
• The audience (humans) may not notice staleness until they hit a problem
• There is no automated validation that docs match code
• Docs live in a separate system (wiki, Notion) disconnected from the codebase
• Nobody owns documentation maintenance as a primary responsibility

The Cost of Rot

• New team members learn wrong patterns from stale docs
• AI agents given stale docs produce code based on outdated assumptions
• Time spent debugging issues caused by following outdated guidance
• Tribal knowledge accumulates in chat, not in any durable format

The Solution: Inverted Documentation Flow

Inverted Flow

Developer modifies a module → Updates the co-located CLAUDE.md in the same PR
                                       ↓
                                Navigation skill describes HOW to explore, not WHAT exists
                                       ↓
                                Plugin bundles skills so agents can load them on demand
                                       ↓
                                Agent enters the module → loads skill → reads live source code
                                       ↓
                                Guidance stays accurate because the source itself is the authority

Why it stays current:

• CLAUDE.md files live *in the codebase*, next to the code they describe
• They are loaded automatically by AI agents when entering a directory
• Skills teach agents *how to explore*, not *what the code currently does* -- so they stay accurate even as implementation details change
• The agent reads current source code, guided by the skill -- the source is the documentation
• Code review can enforce CLAUDE.md updates alongside code changes

Implementation Pattern

Step 1: Each Module Gets a CLAUDE.md

Place a CLAUDE.md file at the root of each significant module, library, or directory.

What goes in a CLAUDE.md:

# {Module Name}

## Purpose
{One paragraph: what this module does and why it exists}

## Entry Points
- `src/index.ts` -- Public API surface. Start here for understanding exports.
- `src/core/engine.ts` -- Core logic. Start here for understanding internals.

## Architecture
{Brief description of how the module is structured internally}

### Key Files
| File | Responsibility |
|------|---------------|
| `src/index.ts` | Public API, re-exports |
| `src/core/engine.ts` | Core processing engine |
| `src/core/types.ts` | Shared type definitions |
| `src/utils/helpers.ts` | Utility functions |

## Conventions
- {Naming conventions specific to this module}
- {Error handling patterns}
- {Testing patterns}

## Dependencies
- `{dependency-a}` -- Used for {purpose}
- `{dependency-b}` -- Used for {purpose}

## Cross-References
- See `../shared/CLAUDE.md` for shared utilities
- See `../api/CLAUDE.md` for the API layer that consumes this module

Key properties of a good CLAUDE.md:

• Hierarchical: Organized as a tree the agent can navigate level by level
• Cross-referenced: Links to related CLAUDE.md files in other modules
• Entry-point focused: Tells the agent *where to start*, not *everything that exists*
• Convention-documenting: Captures patterns the agent should follow when modifying code
• Brief: Under 100 lines. Details live in the code itself.

Step 2: Create Navigation Skills

A navigation skill teaches the agent *how to explore* a library or module. It does not describe the library's current state -- it describes the *process* for understanding it.

Navigation Skill Template:

---
name: {library-name}-navigation
description: >
  Teaches the agent how to navigate and understand the {library-name} library.
  Triggers: "{library-name}", "how does {library-name} work",
  "understand {library-name}".
---

# Navigating {Library Name}

## Quick Orientation
1. Read `{root}/CLAUDE.md` for purpose and entry points
2. Read `{root}/src/index.ts` for the public API surface
3. Read `{root}/src/core/types.ts` for core type definitions

## Understanding the Architecture
1. Start with the entry point identified in CLAUDE.md
2. Trace the main flow: {describe the primary call path}
3. Key abstractions: {list the 3-5 most important interfaces/classes}

## Common Tasks

### Adding a New Feature
1. Define types in `src/core/types.ts`
2. Implement core logic in `src/core/`
3. Export from `src/index.ts`
4. Add tests in `tests/`

### Fixing a Bug
1. Identify the failing test or behavior
2. Trace from the public API inward
3. Core logic is in `src/core/engine.ts`
4. Edge cases are typically in `src/utils/helpers.ts`

### Understanding a Specific Feature
1. Search for the feature name in `src/core/`
2. Read the test file for that feature -- tests document behavior
3. Check CLAUDE.md for any conventions specific to that area

## Anti-Patterns to Avoid
- {Pattern to avoid and why}
- {Pattern to avoid and why}

Why skills work better than static docs:

• The skill tells the agent *what to do*, not what the code is
• The agent then reads the *current* source code to understand what the code is
• The process described in the skill remains valid even as the code changes
• It is a recipe, not a snapshot

Step 3: Package as a Plugin

Bundle related navigation skills into a Claude Code plugin:

{library-name}-docs/
├── plugin.json
└── skills/
    ├── navigation/
    │   └── SKILL.md          # How to navigate the library
    ├── contributing/
    │   └── SKILL.md          # How to contribute to the library
    └── troubleshooting/
        └── SKILL.md          # How to debug common issues

plugin.json:

{
  "name": "{library-name}-docs",
  "description": "Agent-first documentation for {library-name}",
  "version": "1.0.0"
}

Step 4: Agent Loads On Demand

When an agent encounters the library:

1. The plugin's skills appear in the agent's available skill set
2. Agent loads the navigation skill when it needs to work with the library
3. Skill guides the agent to read CLAUDE.md files and explore current source
4. Agent builds understanding from current code, not stale documentation

The Hierarchy: Three Levels of Agent Documentation

Level 1: CLAUDE.md files (per-directory, auto-loaded)
    ↓
Level 2: Navigation skills (per-library, loaded on demand)
    ↓
Level 3: Plugin packages (distributable, installable)

Level 1: CLAUDE.md Files

• Scope: One directory or module
• Loaded: Automatically when the agent enters the directory
• Content: Purpose, entry points, conventions, cross-references
• Maintained by: The team that owns the module
• Update trigger: Code review -- CLAUDE.md changes alongside code changes

Level 2: Navigation Skills

• Scope: One library or subsystem (may span multiple directories)
• Loaded: On demand when the agent needs to work with the library
• Content: Exploration process, common tasks, anti-patterns
• Maintained by: The team or developer who created the skill
• Update trigger: When the library's architecture changes (not every code change)

Level 3: Plugin Packages

• Scope: A distributable collection of skills for a library or framework
• Loaded: Installed into a project, then skills load on demand
• Content: Multiple skills covering navigation, contributing, troubleshooting
• Maintained by: The library maintainers or community
• Update trigger: Major version changes or new skill additions

Designing CLAUDE.md for Machine Consumption

Structure for Machines, Not Narratives

Bad (human-narrative style):

This module was originally created in 2023 to handle user authentication.
Over time, we've added OAuth support, session management, and rate limiting.
The main file you'll want to look at is auth.ts, which contains most of the
logic. There's also a helpers file with some utility functions.

Good (machine-structured style):

# Auth Module

## Purpose
User authentication: login, logout, session management, OAuth, rate limiting.

## Entry Points
- `src/auth.ts` -- Core auth logic (login, logout, verify)
- `src/oauth.ts` -- OAuth provider integrations
- `src/session.ts` -- Session management
- `src/rate-limit.ts` -- Rate limiting middleware

## Key Types
- `AuthUser` in `src/types.ts` -- Authenticated user object
- `Session` in `src/types.ts` -- Session state
- `OAuthProvider` in `src/oauth.ts` -- Provider interface

## Conventions
- All auth functions return `Result<T, AuthError>` -- never throw
- Sessions are stored in Redis -- see `src/session.ts` for connection setup
- Rate limits are per-IP by default -- see `src/rate-limit.ts` for config

Key Differences

CLAUDE.md Hierarchy and Inheritance

CLAUDE.md files are hierarchical -- an agent entering a directory loads the CLAUDE.md from that directory AND all parent directories up to the project root.

project/
├── CLAUDE.md                    # Project-wide conventions (loaded everywhere)
├── src/
│   ├── CLAUDE.md                # Source code conventions (loaded in src/ and below)
│   ├── auth/
│   │   ├── CLAUDE.md            # Auth module specifics (loaded in auth/ and below)
│   │   └── oauth/
│   │       └── CLAUDE.md        # OAuth specifics (loaded only in oauth/)
│   └── api/
│       └── CLAUDE.md            # API module specifics
└── tests/
    └── CLAUDE.md                # Testing conventions

When an agent works in src/auth/oauth/, it loads:

1. project/CLAUDE.md -- project-wide conventions
2. project/src/CLAUDE.md -- source code conventions
3. project/src/auth/CLAUDE.md -- auth module conventions
4. project/src/auth/oauth/CLAUDE.md -- OAuth-specific conventions

Use this hierarchy intentionally:

• Project root: language, build commands, Git conventions, overall architecture
• src/: coding conventions, import patterns, error handling patterns
• Module level: module-specific entry points, types, dependencies
• Sub-module level: only when there are sub-module-specific conventions

Migration Path: From Wiki to Inverted Docs

Phase 1: Add CLAUDE.md Files (1-2 days)

1. Identify the 5-10 most important modules in the codebase
2. Create a CLAUDE.md for each with: purpose, entry points, key files, conventions
3. Add a root CLAUDE.md with project-wide conventions
4. Do not delete the wiki yet -- CLAUDE.md files supplement it initially

Phase 2: Create Navigation Skills (1-2 days)

1. For each major library or subsystem, create a navigation skill
2. Focus on the exploration process, not the current state
3. Package skills into a plugin
4. Test: give the agent a task in each library area and verify it loads the skill

Phase 3: Enforce Co-Location (Ongoing)

1. Add to code review checklist: "Does this change need a CLAUDE.md update?"
2. When new modules are created, require a CLAUDE.md as part of the PR
3. When existing docs are found to be stale, update the CLAUDE.md (not the wiki)
4. Gradually, the CLAUDE.md files become the authoritative source

Phase 4: Deprecate the Wiki (When Ready)

1. Audit: for each wiki page, verify the information exists in CLAUDE.md files or skills
2. Archive the wiki (do not delete -- it may have historical context worth preserving)
3. Redirect documentation questions to: "Read the CLAUDE.md in the relevant directory"

Measuring Documentation Health

Staleness Indicators

Health Metrics

• Coverage: What percentage of significant modules have a CLAUDE.md?
• Freshness: When was each CLAUDE.md last updated relative to its module's last code change?
• Usefulness: Do agents produce better output when CLAUDE.md files are present?
• Completeness: Does each CLAUDE.md have: purpose, entry points, key files, conventions?

Anti-Patterns

1. CLAUDE.md as a Code Dump

Problem: CLAUDE.md lists every file and function in the module. Fix: Focus on entry points and navigation, not exhaustive inventory. The agent can read the directory listing itself.

2. Narrative Prose Instead of Structure

Problem: CLAUDE.md reads like a blog post about the module's history. Fix: Use tables, lists, and labeled sections. Agents parse structure, not stories.

3. Duplicating Code Comments

Problem: CLAUDE.md repeats what is already in code comments and docstrings. Fix: CLAUDE.md should describe the *module-level* view -- architecture, conventions, entry points. Code comments handle the *function-level* view.

4. Never Updating CLAUDE.md

Problem: CLAUDE.md is written once and never touched again. Fix: Make CLAUDE.md updates part of the code review process. If you changed the module's architecture, update the CLAUDE.md.

5. One Giant CLAUDE.md at the Root

Problem: All documentation is in a single root CLAUDE.md file. Fix: Use the hierarchy. Root CLAUDE.md has project-wide conventions; each module has its own CLAUDE.md with module-specific guidance. This mirrors progressive disclosure.

Integration with SDD

Documentation inversion is a natural extension of SDD's context architecture:

The key insight from SDD applies directly: strong documentation enables agents to orient themselves in unfamiliar code and contribute meaningfully without step-by-step human direction.

Cross-References

• context-architecture -- The progressive disclosure pattern that CLAUDE.md hierarchy implements
• methodology -- How documentation inversion fits into the broader Cavekit lifecycle
• spec-writing -- Specs and CLAUDE.md files share the same structural principles
• brownfield-adoption -- When adopting SDD on an existing codebase, CLAUDE.md files are step 1
• impl-tracking -- Implementation tracking documents are a form of inverted documentation