Docgen

Documentation Generator

Overview

Generate or update project documentation from the codebase — READMEs, API references, architecture overviews, and changelogs. Designed to produce accurate, non-padded docs that match the project's existing style.

Core principles:

1. Generate to conversation first. Never write a file without asking. Always: "Want me to write this to [path]?"
2. Preserve manual content. If a doc already exists, identify what's outdated vs. what to keep. Don't blow away prose that was written by hand.
3. Detect existing style. Match the formatting, heading levels, and tone of existing docs in the project.
4. No filler. Omit empty or generic sections rather than padding them with boilerplate.

Invocation Examples

/docgen readme                          # generate or update project README
/docgen readme src/auth/                # README for a subdirectory or package
/docgen api src/routes/                 # API endpoint documentation
/docgen architecture                    # architecture overview from codebase structure
/docgen changelog v0.2.0..v0.3.0       # changelog from git history for a version range

Workflow Order (MANDATORY)

You MUST follow this order. No skipping steps.

Step 1: Parse Input

Determine doc type and scope:

If no argument or unrecognized type: say "Usage: /docgen readme, /docgen api src/routes/, /docgen architecture, /docgen changelog [range]." Then stop.

If api is invoked without a path: say "Please provide a path. Example: /docgen api src/routes/." Then stop.

Step 2: Gather Context

Context to gather varies by doc type:

readme:

• List the project root (or specified subdirectory) to understand structure
• Read package.json, go.mod, pyproject.toml, or equivalent (for name, version, description, dependencies)
• Read entry point files (main, index, cmd/) — top 3-5 most important
• If a CLAUDE.md exists at the project root, read it (especially Architecture and Commands sections)
• Read any existing README to understand what's already there

api:

• List and read the specified directory (routers, handlers, controllers)
• Identify route definitions: look for patterns like router.GET, app.post, @app.route, r.HandleFunc
• Read the top 5-10 route-defining files
• Note authentication patterns, request/response shapes (from types, schemas, or inline validation)

architecture:

• Get the directory tree (2-3 levels deep)
• Read CLAUDE.md if it exists (Architecture section)
• Read the primary entry point(s)
• Identify key packages/modules and their roles from structure + names

changelog:

• Run git log --oneline [range] to get commits
• If no range given, use recent commits since last tag: git log $(git describe --tags --abbrev=0)..HEAD --oneline
• Group commits by conventional commit type (feat, fix, docs, refactor, chore, perf, test)
• Deduplicate: merge commit messages, fixup commits, and WIP commits are filtered out

Step 3: Check Existing Docs

Before generating, check whether the target doc already exists:

• If no existing doc: generate from scratch
• If existing doc found: read it fully, then identify:
  • Sections that are outdated (version numbers, stale API descriptions, removed commands)
  • Sections that look manually written (custom prose, examples, migration notes) — preserve these
  • Sections that can be regenerated cleanly (install commands, dependency lists, endpoint tables)
  • State your preservation plan: "I'll keep the Deployment section as-is and regenerate Install and API sections."

Step 4: Generate Document

Produce the document in the conversation. Structure by type:

readme:

# <project name>

<1-2 sentence description — what it does and who it's for>

## Install
<install command(s)>

## Usage
<primary usage pattern with example>

## Architecture  (omit if trivial)
<brief description of key components>

## Configuration  (omit if none)
<key config options>

## License
<license name>

api:

# API Reference

## Endpoints

| Method | Path | Auth | Description |
|--------|------|------|-------------|
| GET | /api/users | JWT | List users |
...

## <Endpoint Group Name>

### GET /api/users/:id
<description>

**Request**
- Path params: `id` (string) — user UUID

**Response**
- 200: `{ id, name, email, createdAt }`
- 404: `{ error: "not found" }`

architecture:

# Architecture

## Overview
<2-3 sentences: what the system does, how it's structured>

## Components
<table or list of top-level packages/modules with one-line descriptions>

## Data Flow
<how data moves through the system — input → processing → output>

## Key Files
<table: file/path → purpose>

changelog:

## <version or date range>

### Added
- <feat: commits>

### Fixed
- <fix: commits>

### Changed
- <refactor:, perf: commits>

### Internal
- <docs:, chore:, test: commits — summarized, not listed verbatim>

Apply style matching: if the existing project uses ## for top-level sections, match that. If it uses sentence case headings, use sentence case.

Step 5: Present for Review

After outputting the document to the conversation, ask:

Want me to write this to [path]?

If the user says yes: write the file using the Write or Edit tool.

• If updating an existing doc: use Edit for targeted changes, Write only for full replacement.
• Prefer Edit when preserving large sections of existing content.

If the user requests changes: make them in the conversation first, then ask again before writing.

Scope Controls

• Output to conversation first — always ask before writing to disk
• Do not modify files that weren't specified in the invocation
• Preserve manually written prose — only regenerate stale or mechanical sections
• No file writes without explicit user confirmation ("Want me to write this to [path]?")

Common Mistakes