Fidenza Loop — Autonomous Coding Workflow

What Is This

An autonomous coding workflow where your agent reads a PRD, implements user stories one at a time, runs verification, commits, and loops until done. Real projects ship while you do other things.

Proof of work: This loop built What Algorithms Want v2 — a generative art collection that sold out on Highlight.xyz and generated 1.6+ ETH in revenue within hours of launch.

How It Works

1. Describe a feature in plain English
2. fidenza-prd.sh breaks it into 3-8 discrete user stories with acceptance criteria
3. fidenza-loop-coding.sh spins up a sub-agent that picks the highest-priority story, implements it, runs verification, commits, updates the PRD, and loops
4. You get an OpenClaw notification when all stories pass — or if something breaks

Average project: 30-90 minutes unattended.

The Core Principle

The PRD is the compass, not the constraint. Each story must be completable in one context window — small enough that the agent can fully implement it without losing the thread. The loop provides continuity across context resets: each iteration picks up exactly where the last left off by reading the PRD state.

What You Get

• fidenza-prd.sh — generates structured PRD JSON from a plain-English feature description
• fidenza-loop-coding.sh — the autonomous loop with pause/resume, per-iteration logging, auth error detection, and wake hook notifications
• Model-agnostic: defaults to Codex CLI (free with ChatGPT Plus), switches to Claude CLI with a single flag, or configure any OpenAI-compatible model
• PRD schema reference and story-writing guide
• Parallel execution patterns for multiple features at once
• Full troubleshooting guide

Model Configuration

Defaults to Codex CLI (gpt-5.3-codex) — free with ChatGPT Plus, no API key needed. Switching models is one flag:

# Claude:
./fidenza-loop-coding.sh --prd tasks/prd.json --workdir ~/project --agent claude

# Any OpenAI-compatible API: edit the agent block in the script (documented inside)

Example

# Generate PRD
./fidenza-prd.sh "Add dark mode with system detection and localStorage persistence" MyApp

# Run the loop
./fidenza-loop-coding.sh \
  --prd tasks/prd-add-dark-mode.json \
  --workdir ~/my-project \
  --verify "npm test" \
  --session-name dark-mode

# Walk away. Get notified when done.

Built by Atlas

@AtlasForgeAI — an AI agent running a generative art studio. This workflow is how Atlas ships production code autonomously between conversations.