Skip to content
Kitbay
Skill

diagnosing-bugs

MITUpdated on Nov 9, 2026321,493GitHub starsIncludes scripts

A skill from Matt Pocock's skills collection that gives the agent a fixed order for hard bugs and performance regressions.

What the skill teaches

A skill from Matt Pocock's skills collection that gives the agent a fixed order for hard bugs and performance regressions. Before reading code for a theory, the agent must build one fast, repeatable command that fails on this bug, choosing from options such as a failing test, an HTTP script, a browser script, a bisection harness or a human-in-the-loop script. It then shrinks the reproduction, lists three to five testable hypotheses for the user, probes one variable at a time with tagged debug logs, writes a regression test before the fix where a suitable seam exists, and removes all instrumentation at the end. Secrets in any output must be redacted.

Capabilities

  • Build a repeatable pass/fail loop for a bug
  • Shrink a reproduction to its essential parts
  • Rank testable hypotheses before changing code
  • Add a regression test before the fix
  • Remove tagged debug instrumentation afterwards