forward-future/loopy

Most AI prompts tell an agent to do something once and hope for the best. The Loop Library project flips that idea: instead of one-shot instructions, it gives you reusable "loops", playbooks where an agent acts, checks its own work, learns from the result, and repeats until the job is actually done. Each published loop spells out what the agent should accomplish, how to verify the latest attempt worked, what to do with what it learned, and when to stop or ask for help. The result is work you can trust and repeat without rebuilding the workflow every time. The repository has two parts. First, a public website where anyone can browse published loops and copy their prompts, no installation needed. Second, an installable companion called Loopy that plugs into coding agents like Cursor, Claude Code, and Codex. With Loopy installed, your agent gains a guided workflow: it can discover repeated work in your codebase and turn it into a loop, find an existing loop from the catalog, audit and repair a loop you already have, interview you to craft a new one, run a loop in bounded passes and hand back an evidence receipt, debrief completed runs to recommend improvements, save loops to a project file for reuse, or prepare a loop for publication. You just describe what you want done in plain language and Loopy figures out the path. This is useful for anyone who works with AI agents on tasks where the first attempt probably won't be the final answer, fixing production errors, improving test coverage, keeping documentation current, reviewing a product, or speeding up a website. A team lead might save a loop for sweeping docs overnight so any engineer (or agent) can reuse it later. A founder might ask Loopy to analyze their repo for repeated work and turn the strongest pattern into a reliable, bounded workflow. Notably, the project is deliberately conservative about safety. Loopy will not start schedules, change production, publish content, send messages, or take destructive action without explicit approval. Saved loops are treated as untrusted reference data, prompts that don't grant permission to run code or access secrets. If a loop contains secrets, Loopy refuses to save it until you sanitize the prompt. The design philosophy is that loops should be bounded, verifiable, and always hand control back to a person when judgment is needed.