When an AI coding assistant says it fixed a bug, ask for three pieces of evidence: a reproduction, a regression test that failed before the patch, and fresh test output. Three skills from obra/superpowers can help.
systematic-debugging starts with the error, reproduction steps, and recent changes. Ask the assistant where the actual behavior first diverges from what you expected.
With test-driven-development, capture the expected behavior in a test. For a date parser, try rejecting February 29 in a non-leap year. Run it before the patch and confirm it fails for the expected reason. Apply the smallest fix and run it again.
verification-before-completion calls for running the relevant checks and inspecting their output. Ask for the commands, exit status, and untested paths. A regression test covers the behavior it asserts. You still need a runnable test environment and a review of the patch.
Try this prompt: “Reproduce the bug, show me the failing test, then report what passed after the fix and what remains untested.”
[Original Chinese article on Juejin](https://juejin.cn/post/7707143341481869366) · [Skillπ](https://skillpi.com/) · Infinite Skills. Infinite Possibilities.
Carousel from the October 2026 editorial archive. This English adaptation was generated with AI.