The master prompt engineer for overnight autonomous AI runs. Tell it what you want in plain English. It writes the whole run: master prompt, plan, verifier, one command. You go to sleep. The AI ships.
⭐ Star this repo if your AI ships while you sleep · Built by AI Guy Official
Everyone prompts AI coding agents one message at a time, babysitting every step. Then the agent stops after ten minutes, or worse, says it's done when it isn't.
The people getting real results stopped writing prompts. They write loops: a spec that pins the work, a checker that isn't the writer, and a finish line the agent can actually prove. That's how a single goal prompt built an entire company overnight (product, brand, landing page, launch videos) for about 500K orchestrator tokens, and how practitioners run 7, 9, even 15-hour autonomous Claude Code sessions that ship.
Loop Maker packages that entire playbook — Anthropic's long-running-agent research, the Ralph Wiggum loop, spec-driven development, plan-delegate-review orchestration — into one skill. You describe the goal. It engineers the run.
Type loop, answer a few questions (one at a time), and it generates:
| File | What it does |
|---|---|
Master prompt (MISSION.md / BUILD-LOOP.md) |
The agent's full instruction set: mission, guardrails, never-ask autonomy rules, orchestration playbook, definition of done |
feature_list.json |
Every feature with verify steps and "passes": false — the agent can't declare victory early |
loop/PROGRESS.md |
Decision log + honesty buckets (done / blocked / cut) — survives crashes and context limits |
AGENTS.md |
Portable rules file read natively by Codex, Cursor, Gemini, Copilot (+ Claude Code via one line) |
| The launch command | One thing to paste. Then you walk away. |
Drop the loop-maker/ folder into ~/.claude/skills/, restart Claude Code, then type loop + what you want. Full instructions: loop-maker/README.md
One file, commands/loop-maker.md, works everywhere. The command is /loop-maker (not /loop — Claude Code already has a built-in /loop):
| Agent | Put it here | Then type |
|---|---|---|
| Claude Code | ~/.claude/commands/loop-maker.md |
/loop-maker |
| Codex CLI | ~/.codex/prompts/loop-maker.md |
/loop-maker |
| Cursor | .cursor/commands/loop-maker.md |
/loop-maker |
| Antigravity / anything else | anywhere in the repo | "Read commands/loop-maker.md and follow it" |
Then just talk to it: /loop-maker I want to add 3 features to my app — build all of them, test everything end to end, and don't stop until it's done.
How it feels: it scans your repo first (so it never asks what it can see), writes a draft mission prompt instantly from whatever you said, and scores it — Mission Strength: 7/10 — then asks you ONE question at a time (three max) and you watch the score climb. Say "just send it" anytime and it ships the draft with its assumptions logged. At 10/10 you get a Mission Card: codename, run-length estimate, the perfect prompt in one copy block, and the launch line. See you in the morning. 🌙
Three parts, non-negotiable. The Spec pins exactly what to build and what's out of scope. The Verifier is never the writer — tests, screenshots, or a fresh-context skeptic grade the work. The Stop Condition is provable, with a hard turn/time cap.
Plan, delegate, review. The smart model never builds. It orchestrates cheaper worker models — fan-out researchers, tournaments with judge panels, adversarial skeptics, a completeness critic before any phase counts as done. Same results, fraction of the cost.
Never ask, always log. The agent doesn't stop to ask questions at 3 AM. It decides with research, logs the question + answer + why in a Decision Log, and keeps moving. Blocked after three attempts? It ships the strong 80% and tells you exactly what got cut. Blocked is never dressed up as done.
Evidence, not vibes. Every "it works" comes with the test output, the command result, or the screenshot — verified from the source of truth, never a green badge.
The full playbook lives in loop-maker/references/: orchestration patterns · 17 documented failure modes and their guardrails · the five-gate thinking method · launch commands per platform.
Is this safe to run overnight? Every generated loop is scoped to one folder/branch, can't push, deploy, publish, send, or spend, and carries a hard cap. The morning review is part of the method: the loop builds, you stay the boss.
Does it only work for code? No. Build mode handles features, refactors, bug hunts, migrations, content batches. Venture mode runs the full pipeline: hunt for real pain → tournament-pick the winner → design the business → build brand + product → launch videos → red team → packaged recap.
What does an overnight run cost? On a Claude subscription, the plan-delegate-review pattern is the whole trick: the expensive model only orchestrates, cheap models do the building. Scope tight, cap turns, review in the morning.
Join Evolving AI Hub on Skool — the community where we drop every skill, prompt system, and build like this one, free.
Want this level of AI systems installed across your whole business? That's literally what we do — aiguyofficial.com
Loop Maker by AI Guy Official · MIT License · PRs welcome