Dan Trapp

Open-source infrastructure / AI engineering

Claude Code Workflow

An open-source workflow system for running Claude Code with repo rules, specs, PR discipline, and agent boundaries built in.

A published MIT-licensed workflow kit for solo builders, small teams, and engineering orgs that want Claude Code to run inside a repeatable engineering process.

Status

Live OSS

Timeline

Published open-source workflow system

Domain

AI engineering

Why

Open-source infrastructure

View GitHubRead WorkflowRun SetupAsk about this build

Flagship OSS workflow

Spec → Build → PR → Review

Live OSS

Files

6 core

Examples

8

Setup

~5m

Rules, boundaries, and review artifacts for AI-assisted development.

Stack

Languages, services, data sources, and operating pieces behind the build.

CLAUDE.mdAGENTS.mdSPEC_TEMPLATE.mdPR_TEMPLATE.mdWORKFLOW.mdsetup.shMIT License

Code Proof

What The Build Actually Contains

Files

6 core

Examples

8

Setup

~5m

Pattern

Spec→PR

Workflow preview

Spec To PR

01

CLAUDE.md

Repo rules and build conventions

02

AGENTS.md

Agent autonomy boundaries

03

SPEC_TEMPLATE.md

Scope before code

04

PR_TEMPLATE.md

Explain the change

05

WORKFLOW.md

Repeatable loop

06

examples/

Real project starting points

Product proof

Flagship OSS workflow

Spec → Build → PR → Review

Live OSS

Files

6 core

Examples

8

Setup

~5m

Rules, boundaries, and review artifacts for AI-assisted development.

Implementation

Code Behind The Surface

Spec before build

md

The workflow forces scope before Claude starts building, so fast output has something concrete to obey.

# SPEC_TEMPLATE.md

## Problem
What are we solving?

## Scope
What is included and what is not?

## Acceptance Criteria
How will we know this works?

## Human Decisions
Where should Claude stop and ask?

Agent autonomy boundaries

md

AGENTS.md turns implicit judgment into explicit rules for what Claude can do alone and when it must stop.

# AGENTS.md

## Safe To Do Autonomously
- Read files and trace existing patterns
- Make scoped implementation edits
- Run local tests and explain failures

## Always Stop And Ask

- Deleting any file
- Changing the database schema
- Modifying authentication or authorization logic
- Changing payments or billing
- Changing production configuration

Run the setup

sh

setup.sh asks 8 questions and generates a custom CLAUDE.md in about 3-5 minutes.

git clone https://github.com/dantrapp/claude-code-workflow.git
cd claude-code-workflow
chmod +x setup.sh
./setup.sh

Project Logic

Why This Exists

The useful part is the path: what was broken, what I built, and where the leverage showed up.

Mission

What makes Claude Code output inspectable and repeatable before fast AI-assisted coding turns into unreviewed code debt?

Claude Code can write good code, but output varies when the repo rules, scope, autonomy boundaries, and review expectations live only in someone's head. Fast LLM output creates a consistency problem: decisions disappear, review gets harder, and teams inherit code that moved faster than the process around it.

Build

What Had To Work

I built a complete workflow/control layer: CLAUDE.md for codebase rules, AGENTS.md for autonomy boundaries, SPEC_TEMPLATE.md for scope before build, PR_TEMPLATE.md for review discipline after build, WORKFLOW.md for the full process, setup.sh for fast repo setup, and examples for MCP servers, REST APIs, Next.js apps, pSEO, AEO, data pipelines, conversational AI, and solo builders.

Why It Matters

Spec → Build → PR → Review

Turns AI-assisted coding from one-off prompt sessions into a repeatable engineering workflow with planning, review, and human decision points.

Hard Parts

Fast output creates review pressure

Claude Code can produce a lot of code quickly. The workflow slows down only the risky parts: unclear scope, missing rules, undocumented changes, and untested merges.

Process has to stay light

The system is built around a small loop: Spec -> Build -> PR -> Review. The setup script and examples add structure without turning the workflow into heavy ceremony.

Different teams need different defaults

The repo includes examples for MCP servers, REST APIs, Next.js apps, pSEO, AEO, data pipelines, conversational AI, and solo builders so people can start from a pattern that resembles their work.

Decisions

Make CLAUDE.md the repo-level source of truth for Claude Code rules.
Use SPEC_TEMPLATE.md before build and PR_TEMPLATE.md after build.
Use AGENTS.md to separate safe autonomy from human decisions.
Ship setup.sh so a repo-specific CLAUDE.md can be generated in minutes.

Proof

Published OSS workflow system with setup script, reusable templates, and 8 project examples.

Smoke-tested setup flow: clone -> chmod -> ./setup.sh -> generated CLAUDE.md

Includes: CLAUDE.md, AGENTS.md, SPEC_TEMPLATE.md, PR_TEMPLATE.md, WORKFLOW.md, setup.sh

Examples: MCP server, REST API, Next.js app, pSEO, AEO, data pipeline, conversational AI, solo builder

View GitHub

Tell Me About Your Project

Bring Me The Bottleneck.
I’ll Build The Answer.

Tell me what people are trying to do, where the current path breaks, and what kind of useful answer should exist.

Market Gap

Demand exists, but the answer is missing.

Workflow Drag

The work is still too manual, slow, or scattered.

Product Wedge

A small surface could prove the larger opportunity.

LinkedIn

Quick Note