How to use the Verify Before Claiming Done agent skill

What the skill makes the agent do

The skill is built around one rule: no claim of completion without fresh verification evidence. Before the agent says a test passes, a build succeeds or a bug is fixed, it should identify the command that proves the claim, run the full command, read the complete output including the exit code and failure count, and only then state the result.

The skill also lists common failure patterns with what counts as adequate evidence. For example, a passing linter does not prove a build succeeds, and an agent's report of success is not a substitute for checking the version control diff. It also lists warning phrases such as 'should work now' or 'probably' that the agent should treat as a signal to stop and verify.

• Tests: the test command's output must show zero failures, not an earlier run or a belief that it will pass.
• Build: the build command must exit with code 0. A linter passing does not count.
• Bug fixes: the original symptom must be re-tested and shown to pass.
• Delegated work: when a sub-agent reports success, the agent checks the VCS diff and reports the actual state.
• Regression tests: the skill describes a red-green cycle, where the test is seen to fail without the fix and pass with it.

Install the skill from a ZIP file

Download the ZIP for this skill from this page and unzip it. The skill is a folder containing a single file, SKILL.md. Put the folder in one of two places, depending on how you want it used.

For all your projects, place it in ~/.claude/skills/. For a single project, place it in that project's .claude/skills/ folder. The skill's files should sit inside a folder named for the skill, so that the agent finds SKILL.md.

In the Claude apps, you can instead upload the skill under Settings → Capabilities → Skills.

• Personal install: ~/.claude/skills/
• Project install: .claude/skills/ inside your project
• Claude apps: Settings → Capabilities → Skills, then upload

Trigger the skill with natural requests

You do not need a special command. The skill's description tells the agent to use it before claiming success, before committing or creating a pull request, before marking a task done, and before delegating work to a sub-agent. So ordinary requests that involve finishing work are enough to bring it in.

Be specific about the checks you want if your project has more than one. The skill can only verify with commands the project already has, so naming them helps.

• "Fix the failing login test and confirm the full suite passes before you tell me it's done."
• "Before you open the PR, run the build and show me the exit code."
• "The sub-agent says the migration is finished. Check the diff and tell me what actually changed."
• "Make sure the bug is really fixed by re-running the original reproduction steps."

What a good response looks like

With the skill active, a good completion report names the command it ran, quotes the relevant result, and states the conclusion only after that. If the check fails or cannot be run, the agent should say so with evidence rather than softening the claim.

The skill's own examples show the pattern: run the test command, see the pass count, then report that all tests pass. Phrases like 'it should pass' are the thing the skill is designed to block.

• Good: run the test command, see 34 of 34 passing, then say all tests pass.
• Not acceptable: 'It should pass now' or 'the linter is clean, so the build is fine'.

Safety notes: scripts and network

This skill contains guidance only. The package includes no scripts, and it does not use the network. It does not run anything by itself. Any commands that run are the ones your agent executes as part of checking its work, using the project's existing test, lint or build tools.

Because the agent will run your project's commands, the usual precautions for letting an agent execute code in your repository still apply.

• Runs scripts: no
• Uses network: no
• Requires: nothing listed in the skill metadata

Tips for getting the most from it

The skill works best on projects that have clear test, lint and build commands. Without them, the agent has nothing concrete to verify against. Requirement checks are also more useful when the agent has a written plan or checklist to compare against.

Skip the full check cycle for trivial throwaway work, where the extra runs cost more time than they add.

• Make sure your project has documented test, lint and build commands before relying on the skill.
• For larger tasks, ask the agent to write a checklist of requirements and verify each item.
• Ask for the diff to be checked whenever a sub-agent reports success.