Most prompts ask an agent to do a thing once. The insight behind Forward-Future/loopy is that the work worth automating is rarely one-shot, so the unit it packages is not a prompt but a loop: a short playbook with a built-in check, a next step, and an explicit rule for when to stop.
The distinction it draws
The README makes its case with a single before-and-after. A one-shot prompt says "Make this website faster." A loop says: find the slowest page, make one focused improvement, measure it again, keep the change only if it helps, and repeat until every page meets the target or another pass stops helping. Same goal, but the second version tells the agent how to judge its own result and when to quit.
A good loop, by the project's own framing, answers four questions: what the agent is trying to accomplish, how it will know whether the latest attempt worked, what it should do with what it learned, and when to finish or ask for help. That last question gets unusual emphasis. The README is blunt that loops "are not permission for an agent to run forever," and that the best ones are deliberately bounded, with a real check and a moment to hand control back to a person when judgment or approval is needed.
Two parts, only one of which you install
loopy is really two things in one repository, and the README works hard to keep them separate. The first is the Loop Library website, a public catalog where people and agents browse published loops and copy their prompts, with no installation required. The second is Loopy itself, an optional installable skill that guides an agent through working with those loops. As the README puts it, the website is the library and Loopy is a companion way to work with it. You can hand an agent the website without ever installing the skill.
For agents that do not have the skill, the project exposes the catalog directly: a published agent guide, an llms.txt instruction file, and both JSON and plain-text versions of the catalog. That is a practical touch. It treats "an agent reading this without our tooling" as a first-class case rather than an afterthought.
What the skill actually does
Installed, Loopy covers nine paths. Discover inspects a codebase or coding-thread history for repeated work and turns the strongest candidate into a bounded loop. Find searches the live catalog and recommends up to three published loops without running them. Loop Doctor audits a loop and repairs only the material problems. Adapt tailors a loop to your real tools and limits. Craft interviews you one question at a time and builds a loop when the catalog has no good fit. Run executes a loop in bounded passes and returns an evidence-backed receipt. Debrief reviews completed run receipts. Save appends a loop to a project's LOOPS.md. Publish prepares a preview and submits only after you approve the exact text.
Two design choices stand out. Discovery refuses to call work "repeated" until it sees at least two distinct thread occurrences; a code pattern with no run history is labeled a potential loop, not proven recurrence. And LOOPS.md is treated as untrusted reference data, so a saved prompt does not by itself grant permission to run code, deploy, schedule, or send messages. If an accepted prompt contains secrets, Loopy refuses to save it until you supply a sanitized version.
Where it fits
The README states that Loopy checks the live catalog when recommending a loop, and that it does not quietly start schedules, change production, publish content, or send messages on your behalf. Those claims are the authors' own, not something the README verifies, but they read as a coherent posture rather than marketing: the tool wants to be the thing that proposes bounded, checkable workflows and then gets out of the way.
Installation is a single npx skills add Forward-Future/loopy command with per-agent flags for Codex, Cursor, and Claude Code. The repository is MIT licensed. What it does not include, at least in the portion of the README available here, is any benchmark or evidence that agents following these loops converge faster or produce better results than open-ended prompting. The argument is structural and stated plainly, which is refreshing, but it remains an argument rather than a measured result. For teams already running agents against recurring maintenance work, the loop framing and the nine named paths are worth a look on their own terms.
GitHub: https://github.com/Forward-Future/loopy
Curated by Agent Palisade — practical AI for small and mid-sized businesses.
Top comments (0)