SelfSteeringRalph 🔄

An autonomous AI loop that designs itself, then builds and deploys a product.

No PRD. No predefined tasks. Just a goal and one instruction: "criticize yourself."

What Happened

• Input: 31 lines of text
• Output: Live product on Cloudflare Pages
• Iterations: 20
• Time: ~45 minutes
• Human intervention: Minimal (told it about CLI tools)

The loop:

1. Identified its own weaknesses
2. Built its own decision framework
3. Evaluated 8 product candidates
4. Chose AI Image Compressor
5. Defined MVP scope
6. Built the app
7. Deployed to Cloudflare Pages
8. Pushed to GitHub

Live Result

• 🌐 Product: https://squish-image.pages.dev
• 📦 Product Repo: https://github.com/basfenix/squish-image

The Starting Seed

See loop-seed.txt — this is ALL we gave the loop.

# RALPH v2

## Goal
Build a loop capable of autonomously creating and managing an online product that generates at least €1000/month.

## Phase
Loop development. Do not build the product yet—build the machine that builds it.

## The Loop
You are a fresh AI instance with no memory except this file.

1. READ this entire file
2. FIND the next incomplete task (marked [ ])
3. DO it (create files, research, think)
4. UPDATE this file: mark task [x], add new tasks if discovered
5. APPEND a brief learning to the Log section
6. STOP (next instance continues)

## Success Criteria for the Loop
(What makes this loop "ready" to build a product? Discover and add criteria.)
- [ ] ???

## Tasks
- [ ] List 3 weaknesses of this current loop design

## Log

How to Run It

1. Copy loop-seed.txt to loop.txt
2. Run ./run.sh
3. Watch the iterations
4. /quit between iterations (or wait for auto-complete)

Requires Amp CLI with --dangerously-allow-all permissions.

Files in This Repo

├── README.md              # You're here
├── BREAKDOWN.md           # Full iteration-by-iteration analysis
├── loop-seed.txt          # The 31-line starting point
├── loop.txt               # Final state after 20 iterations
├── run.sh                 # The loop runner script
├── decision-framework.md  # Self-created decision process
├── structure.md           # Self-created folder organization
├── research/              # Product evaluation artifacts
├── decisions/             # Product & tech choices
└── specs/                 # MVP specification

The actual product code lives in a separate repo — the loop was smart enough to separate process from output.

Key Insight

"Separating product code from loop artifacts makes the repo clean—no one wants to see decision frameworks in a product repo."

— Iteration 20

The loop understood that messy process ≠ clean product.

License

MIT — do whatever you want with it.

---

Built by watching an AI design itself. 🍿