DEV Community: Owada Tomohiro

wikigen: Auto-generate specification docs from your codebase

Owada Tomohiro — Sun, 15 Mar 2026 00:26:28 +0000

The problem

Documentation gets written once and goes stale. Engineers don't update it. New team members keep asking "where is this API called from?" and "what's this table for?"

I wanted documentation that stays in sync with code — generated directly from the source, not written by hand.

Existing solutions

DeepWiki by Devin generates wiki-style documentation from repositories. The approach — using AI to read code and produce docs — was exactly what I was looking for.

However, for my use case (batch generation across private repos, CI/CD integration), I needed something I could run from the command line without a UI.

The open-source DeepWiki-Open lets you self-host, but requires Docker + Ollama (for embedding) + an LLM backend. That's a lot of infrastructure for generating docs.

The realization

DeepWiki-Open's pipeline is:

clone → embedding → RAG search → LLM generates docs

But Claude Code can already explore codebases. Give it --add-dir and it uses Read, Grep, Glob, and Bash to find and read whatever files it needs. No embedding, no vector DB, no RAG.

clone → claude -p --add-dir ./repo → reads code directly → writes docs

What I built

wikigen is a single-binary CLI that generates GitHub Wiki from source code.

./wikigen owner/repo

That's it. It clones the repo, lets Claude Code explore the code, and outputs GitHub Wiki-compatible Markdown files.

What gets generated

The document structure follows categories from ISO/IEC 12207 (software lifecycle documentation), filtered to what's actually derivable from code:

Factual (directly from code):

System overview, architecture, API specifications
Data models (from migrations, ORM definitions)
Config, environment variables, build/deploy procedures
Test structure, auth flows, error handling

High-confidence inference (from code patterns):

Processing flows (from function call chains)
Security design (from middleware, validation)

Not generated:

Business requirements, risk assessments, SLAs — anything that would be speculation

The prompt explicitly says: "If there's no code evidence, don't write it. Don't even mention that you couldn't find it."

Multi-repo projects

myproject:owner/frontend
myproject:owner/backend
myproject:owner/shared

Multiple repos get merged into one wiki with cross-repository documentation — architecture pages that show how services interact.

Parallel generation

./wikigen -f repos.txt -p 2 -pp 5

-p 2 runs 2 repos in parallel, -pp 5 generates 5 pages per repo simultaneously.

GitHub Actions integration

Wiki auto-updates when you push to main:

- name: Generate wiki
  env:
    CLAUDE_CODE_OAUTH_TOKEN: ${{ secrets.CLAUDE_CODE_OAUTH_TOKEN }}
  run: wikigen -lang en -pp 3 -local . my-project

No clone needed in CI — the checkout action already has the code.

Error handling

Each page auto-retries up to 3 times
./wikigen -retry regenerates only failed pages
Pages save immediately — partial results survive interruptions

Output format

GitHub Wiki-compatible. Push directly to {repo}.wiki.git:

wiki-output/project/
  Home.md
  _Sidebar.md
  System-Architecture.md
  API-Specification.md
  Data-Model.md
  ...

Cross-page links use [Page Title](Page-Filename) format. _Sidebar.md provides navigation.

See a real example: github.com/tomohiro-owada/wikigen/wiki

What went wrong along the way

claude -p output had commentary mixed in. The generated docs would start with "Sure, I'll create the wiki page for you." Fixed by telling Claude to use the Write tool to save files directly, and having Go read the files instead of stdout.

The dialect incident. My Claude Code session was configured to respond in Kyoto dialect. The generated documentation came out saying things like "This API accepts POST requests, ya know." Added "formal technical language only, no dialects" to the prompt.

Prerequisites

Go 1.22+
git (SSH or PAT)
Claude Code CLI, authenticated

git clone https://github.com/tomohiro-owada/wikigen.git
cd wikigen
go build -o wikigen .
./wikigen owner/repo

I Rewrote Google's Gemini CLI in Go - 68x Faster Startup

Owada Tomohiro — Sat, 24 Jan 2026 08:50:45 +0000

TL;DR

Google's official Gemini CLI has ~1 second Node.js startup overhead
I rewrote it in Go → startup is now 0.01 seconds (68x faster)
Reuses auth from official CLI, so your free tier / Workspace quota just works

https://github.com/tomohiro-owada/gmn

Why I Built This

Google's official Gemini CLI is an amazing tool. Rich TUI, seamless Google authentication, excellent MCP support. I loved using it.

But there was one issue for my use case: startup time.

$ time gemini --version
0.22.2
gemini --version  1.00s user 0.24s system 129% cpu 0.951 total

~1 second just to start. That's the Node.js runtime overhead. Fine for interactive use, but painful when you're calling it repeatedly in shell scripts.

The Solution: gmn

So I rewrote the core functionality in Go. The result is gmn (short for gemini-mini):

$ time gmn --version
gmn version 0.2.0
gmn --version  0.00s user 0.00s system 47% cpu 0.014 total

0.014 seconds. 68x faster.

Benchmarks

Metric	gmn	Official CLI	Improvement
Startup	0.01s	0.95s	68x
Binary	5.6MB	~200MB	35x
Runtime	None	Node.js	-

With API response time included:

$ time gmn "hi"
Hello! How can I help you today?
gmn "hi"  0.01s user 0.02s system 0% cpu 3.205 total

$ time gemini -p "hi"
I'm ready to help. What would you like to do?
gemini -p "hi"  2.13s user 0.53s system 24% cpu 10.933 total

Installation

Prerequisites (Important!)

gmn doesn't have its own authentication. You must authenticate once with the official Gemini CLI first:

npm install -g @google/gemini-cli
gemini  # Choose "Login with Google"

gmn reuses credentials from ~/.gemini/. Your free tier quota or Workspace Code Assist quota applies.

Install gmn

Homebrew:

brew install tomohiro-owada/tap/gmn

Go:

go install github.com/tomohiro-owada/gmn@latest

Binary:
Download from Releases

Usage

# Simple prompt
gmn "Explain quantum computing"

# With file context
gmn "Review this code" -f main.go

# Pipe input
cat error.log | gmn "What's wrong?"

# JSON output
gmn "List 3 colors" -o json

# Different model
gmn "Write a poem" -m gemini-2.5-pro

Technical Details

Discovering the API

Initially, I tried using generativelanguage.googleapis.com (the public Gemini API), but got 403 errors due to OAuth scope mismatch.

Reading the official CLI source code, I discovered it actually uses the Code Assist API (cloudcode-pa.googleapis.com). This is an internal Google Cloud API, not publicly documented.

Auth Reuse

The official CLI stores OAuth tokens in ~/.gemini/oauth_creds.json. gmn reads this file and refreshes tokens when needed:

if creds.IsExpired() {
    creds, err = authMgr.RefreshToken(creds)
}

MCP Support

gmn also supports MCP (Model Context Protocol). It reads the same ~/.gemini/settings.json config:

gmn mcp list
gmn mcp call my-server tool-name arg=value

What's NOT Included

gmn is focused on non-interactive use cases:

Interactive/TUI mode → use official CLI
OAuth flow → authenticate with official CLI first
API Key / Vertex AI auth

Conclusion

This is a love letter to Google's official Gemini CLI. I just needed something faster for scripting.

If you use Gemini in shell scripts or automation, give gmn a try:

brew install tomohiro-owada/tap/gmn
gmn "Hello, World!"

https://github.com/tomohiro-owada/gmn

Acknowledgments

Google Gemini CLI — The incredible original
Google Gemini API — The underlying API

Introducing Free RAG for Claude Code — Save Tokens & Time

Owada Tomohiro — Sat, 25 Oct 2025 01:09:22 +0000

TL;DR

Tired of feeding docs to Claude Code every single time?

With a locally running, free RAG tool (DevRag), Claude Code can find the right documents for you via vector search. You no longer need to remember hundreds of filenames or locations.

Completely free: no API, entirely local
Simple setup: ~5 minutes
Fast: token usage cut to 1/40, responses 15× faster
Repository: https://github.com/tomohiro-owada/devrag

Problems When Letting Claude Code Read Documents Directly

1. Wasting context

Claude Code’s context window is limited.

Every time you have it read an entire document, you burn through a huge amount of tokens.

Example:

You: “Check the project’s API authentication scheme.”
Claude reads docs/auth.md (3,000 tokens)
Claude: “We use JWT-based authentication.”

Those 3,000 tokens are now gone from your prompt budget.

Ask something else later → it reads the whole thing again.

2. It’s hard to know which file to look at

As docs accumulate, you don’t know where things are — and neither does Claude.

You: “Tell me about our Redis caching strategy.”
Claude tries:

docs/architecture.md (4,000 tokens)
docs/caching.md (2,000 tokens)
docs/redis.md (doesn’t exist)

But maybe you only needed 200 tokens of docs/caching.md.

In a project with 10–100 documents:

You don’t know where others documented things
You can’t predict filenames
Asking “Where did we write that again?” becomes daily routine

3. Repeated documentation reading

You often refer to the same docs:

Session 1 → docs/auth.md (3,000 tokens)

Session 2 → again (3,000 tokens)

Session 3 → again (3,000 tokens)

Same file, three times, 9,000 tokens.

Because you always read from the beginning, even if you only need a tiny piece.

RAG Solves All of These at Once

How RAG Works

Once at the beginning: vectorize documents and index them
At query time: retrieve only relevant chunks
Claude reads only the necessary parts

Traditional:

Question → Read whole document (3,000 tokens) → Answer

With RAG:

Question → Vector search relevant part (200 tokens) → Answer

This cuts token usage significantly and increases signal-to-noise ratio.

The biggest benefit:

Claude Code can find what you need even if you don’t know filenames.

DevRag — A Simplified RAG for Claude Code

I built DevRag to make context retrieval simpler and faster for Claude Code.

Features

One-binary: no external DB, no Python
Auto model download on first run
MCP integration as a search tool
Fast: startup ~2 s, search <100 ms
Multilingual support (JP/EN)
No vendor lock-in

Setup (~5 minutes)

1. Download binary

# macOS (Apple Silicon)
wget https://github.com/tomohiro-owada/devrag/releases/latest/download/devrag-macos-apple-silicon.tar.gz
tar -xzf devrag-macos-apple-silicon.tar.gz
chmod +x devrag-macos-apple-silicon
sudo mv devrag-macos-apple-silicon /usr/local/bin/devrag

2. Configure Claude Code

Add to ~/.claude.json:

{
  "mcpServers": {
    "devrag": {
      "type": "stdio",
      "command": "/usr/local/bin/devrag"
    }
  }
}

3. Add some documents

mkdir documents
cp your-notes.md documents/

DevRag indexes automatically when launched.

Actual Usage Comparison

Before (No RAG)

You: “What’s our DB migration method?”

Claude reads:

README.md (5,000 tokens)
docs/database.md (4,000 tokens)
docs/setup.md (3,000 tokens)

→ 12,000 tokens, ~30 seconds

Because you’re guessing filenames.

After (With DevRag)

You: “What’s our DB migration method?”

Claude:

Runs vector search
Finds relevant 300-token snippet

Claude:

“Run npm run migrate. For details see docs/database.md:42.”

→ 300 tokens, ~2 seconds

Summary

Directly reading documents means:

❌ Token waste
❌ Hard to find the right file
❌ Repeat full-reads every session

RAG means:

✅ Token usage cut to 1/40
✅ Responses 15× faster
✅ Filename knowledge not required
✅ Setup in ~5 minutes
✅ Entirely local and free

Let Claude Code retrieve what you need automatically using vector search.

Repository

https://github.com/tomohiro-owada/devrag

License: MIT

Feedback: via Issues

Try it out! 🚀

DEV Community: Owada Tomohiro

wikigen: Auto-generate specification docs from your codebase

The problem

Existing solutions

The realization

What I built

What gets generated

Multi-repo projects

Parallel generation

GitHub Actions integration

Error handling

Output format

What went wrong along the way

Prerequisites

Links

I Rewrote Google's Gemini CLI in Go - 68x Faster Startup

TL;DR

Why I Built This

The Solution: gmn

Benchmarks

Installation

Prerequisites (Important!)

Install gmn

Usage

Technical Details

Discovering the API

Auth Reuse

MCP Support

What's NOT Included

Conclusion

Acknowledgments

Introducing Free RAG for Claude Code — Save Tokens & Time

TL;DR

Problems When Letting Claude Code Read Documents Directly

1. Wasting context

2. It’s hard to know which file to look at

3. Repeated documentation reading

RAG Solves All of These at Once

How RAG Works

DevRag — A Simplified RAG for Claude Code

Features

Setup (~5 minutes)

1. Download binary

2. Configure Claude Code

3. Add some documents

Actual Usage Comparison

Before (No RAG)

After (With DevRag)

Summary

Repository