What is the best Claude skill for writing a bug regression test?

Use regression-test-writer. It reproduces the reported failure, identifies the narrowest observable behavior, writes a test that fails before the fix, and verifies that it passes afterward without weakening assertions.

Which Claude skill should I use for a new test suite?

Use test-scaffolder to match the repository's framework, directory layout, fixtures, naming, and commands. Then add integration-test-designer or contract-test-designer when the suite crosses a process or service boundary.

Are coverage and mutation testing interchangeable?

No. Coverage shows which code executed; mutation testing asks whether the assertions detect controlled faults. Use coverage-gap-finder to locate omissions and mutation-test-runner to assess test strength.

Can these skills modify my tests?

Writing skills can create or edit test files when granted write access. Planning and analysis skills can remain read-only. Review the allowed-tools field and the proposed scope before invocation.

9 Best Claude Skills for Software Testing

Start a Claude testing toolkit with regression-test-writer for reproduced bugs, test-scaffolder for repository-native setup, and coverage-gap-finder for prioritization. Add integration, contract, property, mutation, mock-data, or prompt-regression skills according to the failure boundary you need to protect.

The best Claude testing skill depends on the defect you are trying to prevent. Unit tests protect local behavior, integration tests protect collaboration, contract tests protect compatibility, property tests protect invariants, and mutation tests reveal assertions that do not matter. For LLM features, deterministic expected strings are often the wrong tool entirely.

This list maps each testing purpose to a focused Claude skill.

Skill	Best for	Typical artifact	Writes files?
regression-test-writer	A reproduced bug	Minimal failing regression test	Yes
test-scaffolder	Starting tests in an existing repo	Framework-native test files	Yes
integration-test-designer	Real collaborator seams	Integration scenarios and setup	Usually
contract-test-designer	Producer-consumer compatibility	Contract cases	Usually
property-test-designer	Invariants and input spaces	Generators and properties	Yes
mutation-test-runner	Assertion strength	Surviving-mutation report	No by default
coverage-gap-finder	Prioritizing missing tests	Ranked gap report	No
mock-data-factory	Valid reusable fixtures	Data factory	Yes
prompt-regression-tester	LLM behavior changes	Eval cases and comparison	Usually

1. regression-test-writer: lock down a reported bug

regression-test-writer starts from observable failure evidence: a report, reproduction, log, or failing scenario. It locates the narrowest stable boundary, writes a test that fails for the correct reason on the buggy state, and confirms the test passes with the fix.

This ordering matters. A test written only after the code is fixed may pass without ever proving it could detect the regression.

2. test-scaffolder: match the repository

test-scaffolder discovers the existing test runner, naming, directory structure, fixtures, setup hooks, and invocation commands before generating files. Use it when a module has no tests yet or when you are unfamiliar with the repository's conventions.

3. integration-test-designer: protect collaborator seams

integration-test-designer designs scenarios where the risk lives between components: application and database, service and queue, HTTP client and server, or code and filesystem. It helps decide which collaborators must be real and which can remain controlled.

4. contract-test-designer: prevent compatibility breaks

contract-test-designer focuses on the agreement between a producer and its consumers. Use it for APIs, events, schemas, and versioned payloads where either side may deploy independently. It should cover required fields, tolerated additions, invalid states, and version transitions.

5. property-test-designer: explore large input spaces

property-test-designer turns invariants into generated tests. It is strongest for parsers, serializers, transformations, calculations, state machines, and round trips—places where a handful of hand-picked examples cannot represent the input space.

6. mutation-test-runner: test the tests

mutation-test-runner makes controlled changes such as flipping a condition or replacing a return value, then checks whether the suite fails. Surviving mutations identify weak assertions, missing cases, or code that has no meaningful effect.

Run it on a focused package or changed area first; whole-repository mutation runs can be expensive.

7. coverage-gap-finder: prioritize omissions

coverage-gap-finder combines coverage output with code risk. It ranks uncovered error handling, authorization decisions, persistence transitions, and business branches above low-value getters or defensive lines that are difficult to reach.

8. mock-data-factory: produce valid fixtures

mock-data-factory creates reusable builders or factories that encode valid defaults and make important variations explicit. This reduces brittle fixture duplication and prevents tests from accidentally depending on irrelevant fields.

9. prompt-regression-tester: evaluate LLM behavior

prompt-regression-tester handles outputs that vary across runs. It builds representative cases and scores behavioral criteria instead of asserting one exact response. Use it for prompt changes, model upgrades, routing changes, and retrieval pipeline updates.

Recommended testing stacks

For application code, combine regression-test-writer, test-scaffolder, and coverage-gap-finder. For service boundaries, add integration-test-designer and contract-test-designer. For algorithms, add property-test-designer; for an established suite, run mutation-test-runner selectively.

npx agentscamp add skills/regression-test-writer
npx agentscamp add skills/test-scaffolder
npx agentscamp add skills/coverage-gap-finder

Ask Claude to report the exact command it ran, the pre-fix failure, the post-fix result, and any assumptions about fixtures or environment. That evidence makes generated tests reviewable instead of merely plausible.