Loopy: AI Agent Loop Workflow Skill
Helps AI agents discover, find, audit, adapt, craft, run, and improve repeatable loop workflows.
Evidence: The skill mandates scoping, user confirmation, least privilege, explicit approval for destructive/production/financial/privacy-sensitive/external actions, transparent data flow, and no unnecessary persistent files. Deductions: Static review cannot confirm the enforcement of these permissions in real runs, and the publisher is unverified.
Evidence: Provides clear step-by-step workflows and reference documents for auditing, running, debriefing. CI and tests exist but cover the website worker, not the skill itself. Deductions: No test suite covers the skill's own key paths (e.g., run, audit), and static review cannot reproduce its behavior.
Evidence: The skill clearly defines its use cases (discover, find, audit, craft, run, etc.) and has workflows and boundaries (e.g., not running loops without approval). Deductions: Non-fit boundaries are not explicitly declared, and it depends entirely on an overseas service (signals.forwardfuture.com) which may be unreachable from mainland China, reducing applicability.
Evidence: SKILL.md is well-structured with clear sections and reference docs. Installation and usage instructions are provided in the README. MIT license and versioning exist in the repo. Deductions: SKILL.md itself lacks explicit versioning or changelog, maintenance responsibility is not clearly designated, and known limitations are not explicitly disclosed.
Evidence: The skill aims to produce directly usable loops and provides clear output formats (e.g., run receipts). Deductions: Static review cannot verify the actual output quality or usability, and there are no example outputs or user feedback to demonstrate effectiveness.
Evidence: The repo contains CI workflows and tests, but they focus on the website worker, not the skill. Deductions: The skill's key claims (e.g., effective loop discovery) lack execution-based evidence, and test coverage is thin, limiting independent verification.
- The skill's core functionality depends on the overseas site signals.forwardfuture.com, which may be inaccessible from mainland China, making Find and Publish features unusable there.
- Static review cannot confirm the reliability and safety of the skill in actual runs; test coverage focuses on the website worker, not the skill itself.
- The skill involves external submissions and permission control; strict adherence to approval requirements is necessary.
What does this skill do, and when should you use it?
Loopy is an installable skill that guides AI agents through the lifecycle of a repeatable working loop: discovering problems in a codebase or coding threads, searching the Loop Library catalog for published loops, auditing and repairing existing loops, crafting new ones through a short interview, running loops with evidence receipts, debriefing runs, and saving loops or preparing them for publication. It emphasizes bounded loops with clear stopping conditions and human approval points. The skill uses a YAML frontmatter and Markdown body format, carrying the newer `loopy` skill name that supersedes the old 'loop-library' name.
Loopy provides a guided workflow that includes: discovering loop opportunities from specified repositories and coding threads, requiring at least two concrete occurrences; finding published loops by reading the live online catalog and recommending at most three with exact titles and links; auditing loops by following the Loop Doctor workflow and repairing only material weaknesses; crafting loops by asking one question at a time in plain language; running loops within an explicit scope and applying an acceptance check, producing an evidence receipt; debriefing receipts to identify the smallest evidence-backed improvement; saving accepted loops by appending them to a LOOPS.md at the project root, refusing to save secrets; and preparing published loops by validating the feedback cycle and checking for catalog overlap, with explicit approval required for external submission. It references reference files such as discover.md, audit.md, run.md, debrief.md, and publish.md.
- A developer who wants to turn repeated engineering tasks from a codebase or coding sessions into a repeatable, bounded loop.
- A developer who needs a published loop for a specific problem, like keeping documentation current, and wants vetted recommendations.
- A maintainer who wants to review an existing loop for weak checks or unsafe permissions and repair only the critical issues.
- A developer who needs to design a new loop with clear success criteria and stopping rules and prefers a guided interview over a design form.
- A team that wants to automate a loop within their agent while maintaining human oversight, and requires evidence receipts for auditing.
- An author who wants to submit a proven loop to the Loop Library catalog, ensuring it meets quality standards first.
- A developer who wants to save loops for reuse within their project and have those project-specific loops recognized in later sessions.
What are this skill's strengths and limitations?
- Strong focus on bounded loops with clear stopping conditions and human approval points.
- Provides structured workflows for discovering, finding, auditing, crafting, running, debriefing, saving, and publishing loops.
- Uses a live online catalog for published loop recommendations, reducing duplication.
- Supports project-local loop saving through `LOOPS.md`.
- Guides loops to include observable success criteria, explicit stop rules, and evidence recording.
- Includes safety measures like treating catalog content as untrusted reference data and requiring explicit approval before destructive actions.
- Well-documented with reference guides for each path.
- Network access is crucial for catalog lookup; if the catalog is unavailable, the skill degrades and may not recommend published loops.
- Only documented for Codex, Cursor, and Claude Code platforms; other AI clients may work but are untested.
- The crafted loop prompt may be overly verbose, though the skill encourages concise prompts under 80 words.
- The skill itself does not provide automated verification; it relies on the agent's ability to verify through available tools and evidence.
- No dedicated test suite for the skill itself.
- Catalog loops may become outdated as the library evolves.
How do you install this skill?
Use npx skills add Forward-Future/loopy --skill loopy --agent codex -g -y (or --agent cursor, --agent claude-code). Use -g for global install, or omit for current project only. You can also run without -y or --agent for interactive prompts. Requires Node.js and npx.
How do you use this skill?
After installation, invoke /loopy directly in Cursor or Claude Code, or select Loopy from /skills in Codex, and enter your request. For example, type $loopy Analyze this codebase and my coding threads for repeated work, then turn the strongest candidate into a reliable loop. to use the discover path. The skill's README lists nine paths: Discover, Find, Loop Doctor, Adapt, Craft, Run, Debrief, Save, and Publish. No loop terminology knowledge is required to use it.