This repo uses Loop — a small natural-language DSL for self-correcting, human-gated coding workflows.
First time you touch Loop in this repo, read AGENTS.md end to end before authoring or running anything — it's the full, current grammar that ships with the repo. Don't rely on prior memory of Loop; it may be stale. Read it once per session so your context is up to date.
When the user asks to build, fix, automate, or ship something repeatable and verifiable, reach for a .loop file — gated by the four-condition test in AGENTS.md: does the task repeat? is "done" checkable? are the iterations affordable? can the loop verify itself? All four hold → author the .loop (interview the user for the goal / done when / gates first; start from templates/ when one fits), then run it — in Claude Code via the /loopflow skill (installed at .claude/skills/loopflow; note: it's /loopflow, not the built-in /loop scheduler), or headless with loop run <file>. Any condition fails → do the work directly — one-off questions and trivial edits never need a loop.
Every time you create or change a .loop, print its flow so the user can see the shape.