Guide · Agent skill
How to use the AI Loop agent skill (ai-loop): install and run
Updated · written with AI from the project's own documentation and checked against our catalogue data
The ai-loop skill is an agent skill that makes an AI coding agent build a feature in three bounded phases: Spec, Build and Review. It is for developers who want a written specification, a limit on retries and explicit approval points before risky actions. After this guide you will know what the skill does, how to install it, how to ask for it in plain language and what to check before you rely on its output.
AI Loop Development Workflowsickn33/agentic-awesome-skills · MIT Download the skill Quick start
- 1Download and unzip the skill. Download the ai-loop ZIP from this page and unzip it so that SKILL.md sits in an ai-loop folder.
- 2Place it in a skills folder. Move the ai-loop folder to ~/.claude/skills/ for all projects, or to your project's .claude/skills/ folder for one project.
- 3Or upload it in the Claude apps. In the Claude apps, open Settings, go to Capabilities, then Skills, and upload the skill.
- 4Describe a self-contained feature. Ask for a feature with clear scope and success criteria, and say you want a spec-build-review pass.
- 5Answer the spec questions. Reply to the agent's targeted questions about language, edge cases and acceptance criteria so it can write the spec.
- 6Set the iteration limit and approval gates. State how many build-review iterations are allowed and which actions need your approval before work starts.
- 7Check the review results. Read the requirement-by-requirement review and confirm the spec's definition of done is met before accepting the work.
What is an agent skill and what does ai-loop do?
An agent skill is a folder containing a SKILL.md file that gives an AI agent extra instructions for a type of task. When a skill is installed and relevant, the agent follows its process instead of improvising.
The ai-loop skill runs a controlled development loop. The agent first writes a short spec file listing the objective, exact requirements, edge cases and a concrete definition of done. It then builds only what the spec describes. Finally it checks its work requirement by requirement and loops back to fix any gaps, but only within an iteration budget.
- Spec: plan the work and save it to specs/<feature-name>.md without overwriting existing user work.
- Build: implement exactly what the spec says, without adding features or unrelated refactoring.
- Review: compare the code against the spec, list every gap and name the spec item each one fails.
- Approval gates: pause and ask before destructive, production, credentialed or externally visible actions.
Install the ai-loop skill from the ZIP
Download the ZIP for the ai-loop skill from this page. The skill's source is in the sickn33/agentic-awesome-skills repository, under plugins/agentic-awesome-skills-claude/skills/ai-loop, and the skill's only listed file is SKILL.md.
To install it for all of your projects, unzip it into your personal skills folder at ~/.claude/skills/. To install it for a single project, unzip it into that project's .claude/skills/ folder. Make sure the ai-loop folder containing SKILL.md ends up inside the skills folder.
If you use the Claude apps instead, upload the skill under Settings, then Capabilities, then Skills.
- Personal install: ~/.claude/skills/ai-loop/SKILL.md
- Project install: <your-project>/.claude/skills/ai-loop/SKILL.md
- Claude apps: Settings, Capabilities, Skills, then upload the skill.
How to trigger the ai-loop skill with a natural request
You do not need a special command. Describe a self-contained feature and say that you want a complete development pass with a spec and verification. The skill suits requests that have clear success criteria and no unresolved safety or product decisions.
The agent will usually ask a few targeted questions first, such as the language to use or how to handle empty or invalid input. Answer them, and it will write the spec before starting to build.
- Good: I want a utility that calculates mean, median and mode for an array of numbers. Use a spec-build-review loop.
- Good: Build a small date-formatting module with a clear definition of done and at most three review iterations.
- Weaker: Redesign our whole architecture. The skill is built for isolated features, not open-ended refactoring.
Before you start: set the limits
The skill tells the agent to define three things before it begins: the maximum number of build-review iterations, the verification commands or manual checks that count as evidence, and the actions that need explicit approval. You can state these in your request so they appear in the spec.
If the iteration budget runs out while requirements still fail, the agent is supposed to report exactly what remains rather than keep retrying the same fix.
- Iteration budget: a small number, so the loop stops on its own.
- Verification evidence: the test commands or manual checks you accept as proof.
- Approval gates: destructive commands, production changes, external service writes and broad architectural changes.
Example: building a small statistics helper
In the skill's example, a user asks for a utility that calculates mean, median and mode. The agent asks which language to use and how to handle empty arrays and non-number values. After the user answers JavaScript, returning null for empty arrays and ignoring non-numbers, the agent saves specs/stats-calculator.md and then builds the file.
During review, the agent checks each requirement. In the example, the mode function did not filter out non-numbers, so the review marked that item as FAIL. The agent looped back to Build, fixed it and reviewed again until every requirement passed.
Safety notes and limits
The skill itself lists no required tools, does not run scripts and does not use the network. The code it writes is still code you should review and run carefully. The skill's own guidance says to run generated code and tests in a safe, sandboxed environment, to avoid running shell commands from users without checking them, and to keep secrets and credentials out of code and specs.
The review step depends on the agent's self-assessment against the spec. The skill itself says manual review is still recommended for critical systems, and that it is not a replacement for human approval on security-sensitive, production, compliance or externally visible changes.
- Run generated code and tests in a sandbox.
- Never hardcode keys, tokens or credentials in code or specs.
- Treat production deploys, data migrations, payment flows and credential changes as approval-gated work.
- Do not rely on the agent's self-review alone for security-sensitive code.
What people use it for
- Adding a small, isolated utility function with edge cases written into a spec first.
- Building a self-contained module where you want a checklist-style verification pass.
- Adding a feature with a limit on retries so the agent stops and reports what remains.
- Making a feature change where you want the agent to pause before any destructive or externally visible action.
Frequently asked questions
How do I install the ai-loop agent skill?
Download the ZIP from this page and unzip it so the ai-loop folder with SKILL.md is in ~/.claude/skills/ for all projects, or in your project's .claude/skills/ folder for one project. In the Claude apps, you can instead upload it under Settings, Capabilities, Skills.
Does ai-loop require any tools, packages or API keys?
The skill's listing shows no required items, says it does not run scripts and does not use the network. It is a SKILL.md instruction file, so the agent's own environment is what runs any code it writes.
Is the ai-loop skill free to use?
The skill is listed under the MIT licence. The source does not mention pricing, so check the licence file in the repository for the exact terms before reusing it.
Which AI tools does ai-loop work with?
The skill's metadata lists claude, cursor, codex and gemini as tools. The install steps on this page cover the Claude skills folders and the Claude apps, so follow the documentation for your tool if it differs.
What should I do if the review keeps failing?
The skill caps retries with an iteration budget. If requirements still fail when the budget runs out, the agent should report what remains. Make the spec more precise, or split a large feature into smaller loops, as the skill's pitfalls section recommends.
How is ai-loop different from plan-writing or ask-questions skills?
ai-loop covers the full cycle of spec, build and review with iteration limits and approval gates for one bounded feature. The related skill @plan-writing is for more detailed plans on larger projects, and @ask-questions-if-underspecified is for guidelines on interviewing the user.
New agent skills, repos and AI tools — every week
The best new skills and open-source projects, the fastest-growing repos, and one free prompt. Five minutes to read. Unsubscribe anytime.
Our verdict
Use ai-loop if you want an agent to build a self-contained feature from a written spec, with a retry limit and explicit approval points for risky steps. Pick something else for open-ended architectural refactoring, or for security-sensitive, production or compliance work where the agent's self-review cannot replace a human review.
Download the skill