Loopgen — Development skill for Claude Code

loopgen

Prompt Compiler for creating long-running autonomous loops.

It writes the weather, not just the target — a re-readable pressure field that makes the wrong worlds harder to reach while the loop runs.

Your loop died 10 minutes after you went to sleep.

Not because the task was impossible. It blocked on a decision you could have made before it ever fired, or it declared victory on the first green-looking signal.

loopgen writes the part of the prompt that keeps that from happening.

Give it the thing you're trying to do: close a spec, improve the codebase, push a benchmark, walk a frontend, build a vague idea. It classifies the loop, writes the prompt + state + queue files, resolves the decisions that would stall the run, then hands you one `/goal` line. Paste it into Claude Code or Codex and let it run.

The visible output is intentionally boring:

/goal read .loop//PROMPT.md and execute as .


Quick Start

Ask for the loop you actually want. Same compiler, different track:

Track	Ask	What loopgen composes
Product walkthrough	/loopgen "walk the onboarding flow"	A story-shaped loop: keep the product surface contract, reconcile the visible flow with the storyboard.
Backend benchmark	/loopgen "optimize the checkout API load benchmark"	A frontier-shaped benchmark loop: keep pressure accounting, candidate lineage, traces, and metrics.

Both emit the same kind of fixed kickoff:

/goal read .loop/001-onboarding-flow/PROMPT.md and execute as onboarding flow loop.
/goal read .loop/002-checkout-benchmark/PROMPT.md and execute as checkout API load benchmark loop.

The stable file shape is the point. Product walkthroughs keep the story surface:

.loop/001-onboarding-flow/PROMPT.md
.loop/001-onboarding-flow/STATE.md
.loop/001-onboarding-flow/