How to use the Verification Before Completion agent skill

What is an agent skill and what does this one do?

An agent skill is a folder containing a SKILL.md file of instructions that an AI agent can load when a task matches the skill's description. This skill is guidance text, not a program: it shapes how the agent behaves but does not run tests or tools on its own.

The skill's core principle is 'evidence before claims'. Before the agent says a verifiable result holds, such as 'tests pass', 'build succeeds' or 'bug fixed', it should have evidence for that result on the candidate and environment it names. It should also say whether it ran the check itself or relied on someone else's result.

• Requires a test run, build exit code or checked artifact before a result claim, not just a 'should pass' guess.
• Tells the agent to re-run checks when the relevant inputs change.
• Matches the depth of checking to the stakes, so a one-line copy tweak does not need a full test suite.
• Lists common failure patterns, such as treating a linter pass as proof that the code compiles.
• Warns that a passing check can still prove nothing, for example when its output looks like success on failure too.

Install the Verification Before Completion skill

Download the skill as a ZIP from the directory page. You can then add it in one of two ways, depending on how you use Claude.

For Claude Code or a local setup, unzip the download into your personal skills folder at ~/.claude/skills/, or into a project's .claude/skills/ folder if you want the skill only for that project. For the Claude apps, upload the ZIP under Settings, then Capabilities, then Skills.

• Personal use across projects: unzip into ~/.claude/skills/.
• One project only: unzip into the project's .claude/skills/ folder.
• Claude apps: upload the ZIP under Settings → Capabilities → Skills.

Trigger the skill with natural requests

The skill's description says it applies when the agent is about to claim work is complete, fixed or passing, or before committing or creating pull requests. You do not need a special command. Ask for the situation the skill covers and state what evidence you want.

Name the check you care about. The skill works with whatever checks your project already has, so tell the agent which test command, build or artifact should back the claim.

• 'Fix the login bug, then confirm the original failing test now passes before you say it is fixed.'
• 'Before you commit, run the project's tests and tell me which ones you ran yourself.'
• 'Before you open the PR, check the build exits 0 and list anything you did not verify.'
• 'Review the sub-agent's report and check the artifact it names before we mark this done.'

Read the agent's report to see what was checked

A useful result from this skill is a report that separates what the agent ran from what it took on trust. The skill asks the agent to state whether it ran a check or inspected another owner's evidence, and to name the scope of any partial check.

When an agent delegates work or reports success, the skill expects the agent to inspect the artifact and the evidence it names rather than pass on the report alone. Read the report with that in mind: a claim like 'tests pass' should point to the candidate it ran against.

• Look for whether the agent says 'I ran' or 'the sub-agent reported'.
• Check that the tested candidate is the one you are about to ship.
• Note any gaps the agent names, such as untested platforms or skipped checks.

Know the limits and safety notes

The skill declares no required tools or packages, does not run scripts, and does not use the network. Its bundled files are SKILL.md, references/performance-claims.md and LICENSE.txt.

Whether any check actually runs depends on your project and the agent's environment. If the agent has no test runner, no access to the build, or no permission to run commands, it cannot produce the evidence the skill asks for. In that case it should say so rather than claim success.

The skill also scales the work down. For exploratory or throwaway tasks, full verification can slow you down without benefit, and the skill itself says to match checks to the stakes.

• No scripts, no network access, no listed requirements.
• Checks run only if the agent's environment allows them.
• Not suited to creative writing or opinions, where there is no test or build result to check.

Use the performance-claims reference for speed claims

When the agent claims that a change makes something faster, lighter or more responsive, the skill points it to references/performance-claims.md. That file asks for claims to be tied to the versions, fixture and workload measured, and for both time and correct work completed to be counted.

The reference also asks for repetitions and variability to be reported, and for honest negative results to be kept. It is meant for performance claims only, not for every change.

• Bind the claim to the versions, fixture and workload you measured.
• Report repetitions, variability and limits.
• Count failures as well as time, since a speedup from doing less work is not a real speedup.