Skill Test Lab -- Break Your Skill Before Buyers Do

A valid SKILL.md can still produce terrible behavior.

Skill Test Lab evaluates what an agent actually does when your instructions meet realistic requests. It tests the happy path, near-misses, missing inputs, unavailable dependencies, and the permission traps that turn a useful assistant into a liability.

What you get

• A compact test matrix tied to the skill's promised outcome and boundaries
• Normal, near-match, underspecified, failure, and authorization-trap cases
• Isolated execution that stubs external mutations unless a real test is explicitly approved
• Evidence for actions attempted, artifacts produced, and observable results
• Pass, partial, fail, and not-run classifications with causes separated
• A ready, ready_with_limits, or not_ready verdict
• The smallest revision supported by failed behavior, plus focused regression cases
• A record of what remains untested

This product tests behavior, not whether the response repeated expected words. It never publishes the skill, installs it globally, sends messages, spends money, or touches production as part of evaluation.

Use it before a marketplace launch, after a substantial skill revision, or when customers report that a skill follows its prose but misses its job. It pairs naturally with Release Candidate Audit when the package around the skill also needs inspection.