DEV Community

Zac
Zac

Posted on

18 production CLAUDE.md templates (organized by stack and team size)

#claudecode #ai #productivity #tutorial

Most CLAUDE.md files I've seen on GitHub cover code style. Almost none cover what actually causes autonomous runs to break.

I spent 72 hours running Claude Code as an autonomous agent. Here's the template library I wish existed when I started.

What's in the repo

By stack:

  • Next.js + Supabase
  • FastAPI + PostgreSQL
  • React + Vite (TypeScript)
  • Express.js API
  • Django REST Framework
  • Ruby on Rails
  • SvelteKit
  • Go API
  • Python CLI
  • TypeScript library / npm package

By team size:

  • Solo indie hacker
  • Startup team (3-15 people)
  • Enterprise (50+ people)

By use case:

  • Autonomous agent runs (long sessions, context management)
  • Data pipeline / ETL
  • React Native (Expo)
  • Docker / DevOps
  • Monorepo (Turborepo / Nx)

What these cover that most CLAUDE.md files don't

Session recovery — after compaction, the agent re-reads CLAUDE.md but loses conversation context. Most files have nothing about recovery. Every template in this repo has:

## Recovery
After compaction or restart: read tasks/current-task.md
Enter fullscreen mode Exit fullscreen mode

One line. Saves 15 minutes every time context resets.

Error recovery patterns — what to do when a tool call fails. Without this, you get infinite retry loops at 3am.

Negative constraints — "never git add -A" prevents more problems than "be careful with git."

Rate limit documentation — explicit handling for every external API the agent calls.

Approval gates — what requires a human check before proceeding.

How to use

  1. Clone the repo
  2. Copy the template closest to your stack
  3. Fill in the [placeholder] sections
  4. Update whenever the agent does something unexpected 
git clone https://github.com/seankim-android/claude-md-templates
cp claude-md-templates/stacks/nextjs-supabase.md your-project/CLAUDE.md
Enter fullscreen mode Exit fullscreen mode

The philosophy

The most effective CLAUDE.md is one you've updated 20 times based on actual agent behavior. These templates are starting points — not finished documents.

Each one is built from watching Claude Code fail in specific ways in that stack. The rules exist because something broke.

Repo: github.com/seankim-android/claude-md-templates

Built during a live experiment at builtbyzac.com. More context management patterns in the Agent Harness ($29).

Top comments (0)