How to use the Test Value Auditor agent skill

What is an agent skill and what does test-value make the agent do?

An agent skill is a folder of instructions and reference files that an AI agent loads when a task matches the skill's description. The test-value skill is built around one question: what real failure would a test catch, and what would a user lose if that failure happened?

When installed, the agent reviews existing tests or picks which tests to add, change, merge or remove for the code it is working on. It recommends the cheapest check that still protects that risk, and it helps diagnose a failing test by separating a real product bug from an outdated assertion, an environment problem or an unstable test.

• Audit existing tests for cost, false alarms and protection within the scope you request. An audit is read-only; an optimization request permits scoped changes.
• Choose tests during implementation or before a commit, limited to the behavior being changed.
• Diagnose a failing check and keep a requested fix narrow.

How do you install the test-value skill?

Download the ZIP file from this page and unzip it. The skill folder should contain agents/openai.yaml, references/audit.md, references/decision-examples.md and SKILL.md.

For a personal install, unzip it into ~/.claude/skills/. For a single project, unzip it into the project's .claude/skills/ folder so it applies only to that repository.

You can also upload the skill in the Claude apps under Settings → Capabilities → Skills.

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

How do you trigger the skill with natural requests?

You do not need a special command. Describe the testing task in plain language and mention the code or tests involved. The agent matches your request against the skill's description.

Requests that fit the skill include auditing slow or flaky tests, choosing tests while implementing a change, and reviewing test changes before a commit or pull request. Keep the scope clear, because the skill is designed to stay limited to affected behavior unless you ask for a broader audit.

• Audit the slow tests in the payments module and tell me which ones protect real behavior.
• I changed the export formatter. Which existing tests should I add, adapt or remove before I commit?
• This test fails after my refactor. Is it a real bug, an outdated assertion, an environment problem or a flaky test?
• Review the test changes in this PR for assertions that just copy what the implementation outputs.

What does the agent check before recommending a test change?

For tests being added or changed, the skill asks three questions: which concrete mistake would fail the test and what loss it would cause; whether an equivalent refactor or a better valid result would also fail; and whether an existing test, type check or other check already catches the same mistake.

The skill prefers observable output, state, data integrity and failure behavior. It also notes that coverage percentages and green runs are evidence, not substitutes for product judgment. When an AI wrote both the implementation and the tests, it checks whether they share the same mistaken assumption.

• Name the failure each test would catch.
• Check whether a valid alternative would wrongly fail the test.
• Look for an existing check that already covers the same mistake.
• Confirm that the intended tests actually ran, including filters and skips.

What does the skill need and what are the safety notes?

The skill lists no required prerequisites. It does not run scripts and does not use the network, according to its metadata. Your agent still works inside the permissions and project tools you have given it, so review its proposed changes before accepting them.

The skill says reviewing before a commit does not authorize committing or pushing. If you want the agent to commit or push, ask for it explicitly. The skill also says destructive probes should run in temporary copies or controlled fixtures, preserving existing user changes.

• Requires: nothing listed
• Runs scripts: no
• Uses network: no
• It does not authorize commits or pushes by itself.
• Destructive experiments belong in temporary copies or controlled fixtures.

What tips help you get better results?

Give the agent a clear boundary, such as one module, one changed file or one failing test. The skill is designed to avoid repository-wide audits on small edits, so a bounded request gets a faster and more focused answer.

Ask for measurements when you care about speed. The skill says speedups should be claimed only from measurements, and it distinguishes case time, file time, a parallel pipeline's critical path and total compute. Ask the agent to narrow failing runs for diagnosis instead of rerunning the full suite each time.

• Name the specific module, test file or behavior in scope.
• Ask for measured timings rather than estimates.
• For a failing test, ask for the diagnosis first, then a narrow fix.
• Keep project-specific conventions in your project guidance, not in the skill.