Agent debugging workflow
Skill Amey-Thakur/AI-SKILLS/skills/ai-coding-tools/agent-debugging-workflow
Work with an agent on a bug by supplying evidence and letting it form hypotheses, rather than asking it to guess from a description. Use when debugging with an agent's help.From its SKILL.md
npx -y skills add Amey-Thakur/AI-SKILLS --skill agent-debugging-workflowAssembled from the repository path, not quoted from the project. Check it against their README if it does not work.
One thing to look at
- 5 stars5 stars. Stars are a popularity signal and not a quality one, but at this level it is likely that nobody has read this closely except its author, and you would be relying on your own review.
SKILL.md
1.8 KB, as published. Nobody here has run it
Agent debugging workflow
Agents debug well with evidence and badly with descriptions. The common failure is describing the symptom and receiving plausible speculation, when the actual error output would have been decisive.
Method
- Give the raw evidence first. Error message, stack trace, logs, and the exact reproduction steps, rather than your summary of them (see stack-trace-reading).
- Withhold your theory initially. Stating your suspicion anchors the agent on it, and independent hypotheses are more useful (see debugging).
- Ask for hypotheses ranked by likelihood. Several candidates with a way to distinguish them beats one confident answer.
- Let it instrument rather than guess. Adding logging or a targeted test to narrow the cause is faster than reasoning about possibilities (see binary-search-debugging).
- Confirm the mechanism before accepting a fix. A change that makes the symptom disappear without an explained cause is not a fix (see root-cause-analysis).
- Provide the failing test if one exists. A reproduction converts debugging from discussion into verification.
- Record the cause when found. The explanation belongs in the commit or an issue, since the next occurrence will look unfamiliar (see incident-postmortem).
Boundaries
An agent cannot observe your running system and works only from what you show it. Intermittent and environment-specific bugs are the hardest to delegate. A fix that passes tests may still be wrong about the cause (see test-flakiness-budget).