Skip to content

Repository files navigation

How I got non-technical PMs to ship production apps with AI

License: Unlicense

Drop this into any project folder and start building real software with Claude Code, Cursor, or Codex. No coding experience required.

This isn't a framework. It's a set of config files and slash commands that turn your AI coding tool into a disciplined development partner -- one that plans before it codes, reviews its own work, and keeps a paper trail.

Based on the workflow Zevi Arnovitz (PM at Meta) used to build StudyMate -- a revenue-generating app he shipped solo without writing code. Adapted here for multi-tool support.

v2 adds artifact persistence, small batch constraints, and work mode routing. See What's New in v2.


What you get

  • 3 config files -- works with Claude Code, Cursor, and Codex simultaneously
  • 8 slash commands -- enforced workflow from idea to shipped feature
  • Cross-model review -- different AIs validate each other's work
  • Progress tracking -- markdown plans with status markers
  • First principles guide -- the reasoning behind the system, so you can build your own

The workflow

Full Build (new features, refactors, anything non-trivial)

/create-issue  →  Capture idea or bug
/explore       →  Analyze the problem, save findings to file
/create-plan   →  Generate implementation plan (links to exploration)
/execute       →  Build step-by-step (~20-50 lines per task)
/review        →  Self-review for bugs (appended to plan)
/peer-review   →  Cross-model validation (appended to plan)
/document      →  Update architecture docs
/learn         →  Understand complex concepts (use anytime)

Quick Fix (< 20 lines, no new files or data changes)

Describe fix → Code it → /review → Ship

See docs/WORKFLOW.md for detailed decision criteria.


Quick start

# 1. Copy template to your project
cp -r project-kit/ ~/projects/my-new-project/
cd ~/projects/my-new-project/

# 2. Customize CLAUDE.md (replace [PLACEHOLDERS])
vim CLAUDE.md

# 3. Open in your editor
cursor .   # or: code .

# 4. Start building
# Type /create-issue to capture your first feature

Read SETUP.md for complete instructions.


What's New in v2

v2 was shaped by a first-principles analysis of the workflow. Three gaps were identified and closed:

1. Artifact persistence -- every phase leaves a trace

Before: /explore findings and /review results lived only in the chat session. Close the window, lose the reasoning.

Now:

  • /explore saves findings to plans/YYYY-MM-DD-[slug]-exploration.md
  • /review appends its summary to the plan file under ## Review Log
  • /peer-review appends validated findings to the same review log
  • /create-plan links back to the exploration file

The plan file becomes the complete record: what was explored, what was built, and what was reviewed.

2. Small batch constraints -- right-sized tasks

Before: No limit on task size. The AI could generate a "Build the entire checkout flow" task with 400 lines.

Now:

  • Each task should produce roughly 20-50 lines of code changes
  • Tasks larger than ~50 lines get broken into subtasks
  • Phases with more than 5 tasks get split into two phases
  • Plan quality checklist enforces this before execution begins

3. Work mode routing -- right process for the job

Before: One workflow for everything. A CSS color change went through 8 phases.

Now: Two modes, determined before work starts:

Mode When Workflow
Quick Fix < 20 lines, no new files, no data/auth changes Code it → /review → Ship
Full Build Everything else All 8 phases

When in doubt, use Full Build. The cost of over-planning is minutes; the cost of under-planning is hours.


Why this works

The problem: AI coding tools are eager to write code without understanding the problem. That leads to bugs, rework, and frustration.

The fix: Force planning before coding using slash commands.

What makes it different:

  1. Your project folder IS the CTO -- no separate ChatGPT window needed
  2. Multi-tool support -- Claude Code, Cursor, and Codex all read the same config
  3. Cross-model review -- different models catch different bugs
  4. "Less context" framing -- primary model validates peer findings (prevents false positives)
  5. Artifact persistence -- exploration, review, and peer review results survive across sessions
  6. Work mode routing -- right amount of process for the size of the change
  7. Small batch enforcement -- tasks capped at ~50 lines to keep changes reviewable

File structure

your-project/
├── CLAUDE.md              ← CTO system prompt (main config)
├── .cursorrules           ← Cursor config (points to CLAUDE.md)
├── AGENTS.md              ← Codex/OpenCode config (points to CLAUDE.md)
├── FIRST-PRINCIPLES.md    ← The reasoning behind the system
├── .claude/commands/      ← Slash commands (8 files)
├── .cursor/commands/      ← Same commands for Cursor (keep in sync)
├── docs/
│   ├── ARCHITECTURE.md    ← System design (template -- customize)
│   ├── TECH-STACK.md      ← Technology choices (template -- customize)
│   ├── WORKFLOW.md        ← How we work
│   ├── decisions/         ← ADRs (created during use)
│   └── backlog/           ← Issues (created during use)
├── plans/                 ← Exploration + execution plans (created during use)
├── SETUP.md               ← Setup instructions
└── README.md              ← This file

Note: The docs/ARCHITECTURE.md and docs/TECH-STACK.md files are templates with placeholders. You'll need to customize them for your specific project -- they're meant to be filled in as you build, not used as-is.


Who this is for

  • Non-technical PMs who want to build products with AI
  • Solo founders building MVPs
  • Junior engineers learning to build with AI
  • Anyone shipping side projects with AI tools

For multi-person teams: This template is built for solo builders. If you need RFCs, handoffs, multi-model validation pipelines, and tiered process scaling, see agentic-ai-dev-team-setup.


Requirements

Tools (at least one):

AI subscriptions (recommended):

  • Claude Pro ($20/mo) -- for Claude Code
  • ChatGPT Plus ($20/mo) -- for Codex (optional)
  • Cursor Pro ($20/mo) -- for Cursor Composer (optional)

Total cost: $20-60/mo depending on which tools you use.


Tech stack agnostic

This works with any stack -- React, Vue, Svelte, Next.js, Python, Go, whatever. Just customize CLAUDE.md and docs/TECH-STACK.md for your choices.


Documentation

File Purpose
SETUP.md Complete setup guide (start here)
CLAUDE.md CTO system prompt (customize this)
docs/WORKFLOW.md Detailed workflow guide
docs/ARCHITECTURE.md System design template
docs/TECH-STACK.md Technology choices template
FIRST-PRINCIPLES.md The reasoning behind the system

Build your own

This template encodes a specific set of opinions. If you want to build your own AI development workflow from scratch, start with FIRST-PRINCIPLES.md -- it distills 15 principles from elite engineering teams (Google SRE, DORA, Linear, NASA KM) into a reference you can use to design your own system.

The principles are tool-agnostic and process-agnostic. They describe what matters without prescribing how to implement it.


Credits

Inspired by: Zevi Arnovitz (Meta PM) -- original workflow creator, shared on Lenny's Podcast

Adapted by: Benison Joseph


License

Unlicense -- public domain. Use however you want. No attribution required.


Contributing

This is a template, not a framework -- so contributions look a bit different:

  • Bug fixes (broken links, typos, incorrect instructions): PRs welcome
  • New slash commands or workflow improvements: Open an issue first to discuss
  • Adaptations for your stack: Fork it -- that's what it's for

No formal process. Keep it simple.

About

Inspired by Zevi's workflow on Lenny's podcast - project-level configuration to start your product dev

Resources

Contributing

Stars

1 star

Watchers

0 watching

Forks

Releases

Packages

Contributors