Guide · Agent skill
How to use the create-verification-skill agent skill
Updated · written with AI from the project's own documentation and checked against our catalogue data
create-verification-skill is an agent skill that writes another skill for one specific project. The generated skill launches your app, checks it is the right instance, drives one feature the way a user would, saves evidence, and cleans up its own processes. This guide is for developers who want a repeatable way for an AI agent to prove that a web UI, CLI or service behaves correctly, and it explains how to install the skill, trigger it and review what it produces.
Create Verification Skillpaperclipai/paperclip · MIT Download the skill Quick start
- 1Download the ZIP. Download the create-verification-skill ZIP from the skill page on this site.
- 2Install the skill. Unzip it into ~/.claude/skills/, into your project's .claude/skills/ folder, or upload it under Settings → Capabilities → Skills in the Claude apps.
- 3Check the requirements. Confirm you have a Git checkout of the repository, a working local dev or start command, and the browser, tmux/PTY or curl tooling your app needs.
- 4Ask for the skill. Run /create-verification-skill or ask for a control skill for the repo, naming the app and the surface you care about.
- 5Review the output. Read the generated verify-<app> skill and its scripts before running them, and confirm the isolation and cleanup steps.
- 6Run it once. Let the agent run the generated skill end to end on one feature, then check that the evidence remains after cleanup.
What an agent skill is and what this one makes the agent do
An agent skill is a folder with a SKILL.md file that tells an AI agent when to use it and how to carry out a task. This skill tells the agent to interview your repository, then generate a project-local verification skill, normally at .agents/skills/verify-<app>/.
The generated skill is written for the next agent, which will read it cold. It launches the app, runs a doctor check to confirm the instance is worth driving, exercises one feature through a browser, terminal or HTTP calls, and records screenshots, logs, responses or exit codes as proof.
- Surface: the skill identifies what a user touches, such as a web UI, CLI/TUI, desktop app, API, mobile app or library.
- Evidence: proof artifacts are kept after teardown, in a location the generated skill names.
- Maintenance: for Paperclip, the skill reads and reports drift in the feature and agent access maps; other projects get a features/ folder with one file per feature.
Install the create-verification-skill skill
Download the ZIP from the skill page on this site. Unzip it into ~/.claude/skills/ to make it available to you across projects, or into a project's .claude/skills/ folder to share it only with that repository.
If you use the Claude apps, you can instead upload the skill under Settings → Capabilities → Skills.
- Personal install: unzip into ~/.claude/skills/.
- Project install: unzip into the project's .claude/skills/ folder.
- Claude apps: upload under Settings → Capabilities → Skills.
Trigger the skill with a natural request
Ask for it directly with the slash command, or describe what you need in plain language. The skill's description names these triggers: the /create-verification-skill command, a request such as make a control skill for this repo, or a project that has no scripted way to prove UI, CLI or service behavior.
A useful request names the app and the surface you care about, for example a web dashboard or a CLI command, so the agent knows which interface to interview first.
- Use the slash command: /create-verification-skill
- Use a plain request: make a control skill for this repo
- Be specific: say whether the primary surface is the browser, the terminal or an HTTP service.
/create-verification-skillWhat you need before you start
The skill works from a checkout of the target repository, and the app's own dev or start commands must already work locally. The skill tells the agent to fix a broken base first, or report the problem precisely, because a skill written against a broken startup teaches wrong steps.
Browser automation (such as Playwright or CDP), tmux or PTY tooling, or curl are needed depending on the app surface. The generated skill runs scripts and uses the network. For Paperclip itself, it also needs the sibling maps directory and Node.js to run scripts/check-maps.mjs.
- A Git checkout of the target repository.
- The app's dev or start command runs locally.
- Browser automation, tmux/PTY or curl, depending on the surface.
- For Paperclip: the sibling maps directory and Node.js for scripts/check-maps.mjs.
node scripts/check-maps.mjsReview the generated skill and run it once
The skill asks the agent to run the generated skill end to end once before handing it over: launch, doctor, drive one mapped feature, capture evidence and clean up. It then checks that the evidence still exists after cleanup. A generated skill that has never run is a draft.
Read the generated scripts before you run them. The output runs commands and touches processes on your machine, so you should be able to see exactly what it does.
- Confirm the launch command and the readiness signal match your app.
- Check that the doctor step verifies the right port, build and data directory.
- Confirm cleanup kills only processes the run started and leaves the evidence folder intact.
- Run the generated cleanup after any failed attempt so no processes or ports are left behind.
Safety notes: isolation, network and scripts
The skill tells the agent to isolate instances with separate ports, data directories or profiles where possible. Do not point it at a shared or production instance. Even with isolation, driving a live user session can corrupt data.
Dry-run or test modes are not always safe. The skill asks the agent to verify what a dry-run skips by observing files, network activity and git refs rather than trusting the name, since some dry-runs still touch the network or open a browser. Cleanup should never kill processes by name; it should only stop what the run started.
- Use disposable data directories and separate ports.
- Never drive an instance the verification run did not start.
- Verify dry-run behavior by observation, not by its name.
- Review generated scripts before running them.
What people use it for
- Proving a web dashboard's login and settings screens still work after a refactor, with screenshots saved as evidence.
- Checking that a CLI command creates the expected files and returns the right exit code.
- Confirming an HTTP service's endpoints return the expected responses on a disposable local instance.
- Keeping the Paperclip feature and agent access maps honest as the app changes.
Frequently asked questions
How do I install the create-verification-skill skill?
Download the ZIP from the skill page on this site and unzip it into ~/.claude/skills/ for personal use or into a project's .claude/skills/ folder. In the Claude apps you can instead upload it under Settings → Capabilities → Skills.
What do I need to run create-verification-skill?
You need a Git checkout of the target repository and a working local dev or start command for the app. Depending on the surface you also need browser automation such as Playwright or CDP, tmux or PTY tooling, or curl. For Paperclip, you also need the sibling maps directory and Node.js for scripts/check-maps.mjs.
Is create-verification-skill free to use?
The skill is listed under the MIT licence. Its origin and attribution are documented in its NOTICE.md file.
Is it safe to let the agent run the generated scripts?
The generated skill runs commands, uses the network and touches processes on your machine. Review the scripts before running them and only drive isolated instances with disposable data. Do not run it against a shared or production instance.
What if the generated skill does not show up or does not start?
Check that the skill folder contains SKILL.md with its YAML frontmatter, because without frontmatter the skill never registers. If the generated app launch fails, fix the app's own dev or start command first, since the skill assumes that command works.
How is this different from a normal test suite?
A normal test suite is the better choice when one already exists. This skill is for projects that lack a scripted way for an agent to prove UI, CLI or service behavior, and it produces a generated skill plus a feature map that you then have to maintain.
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 this skill if you want an AI agent to run repeatable, scripted checks against a web UI, CLI or service in a project that has no automated end-to-end check. Skip it if your project already has a working test suite, if you cannot review generated scripts, or if you would have to run it against a shared or production instance.
Download the skill