Agent Self Audit
Use this skill as a pre-final or pre-commit self-check for non-trivial work. Its job is to make the agent name what could be wrong with its own result before presenting it as done.
Workflow
- Reconstruct the user intent from the first user request plus any later corrections.
- Score the work against seven dimensions:
- intent satisfaction
- functional correctness
- visual/behavioral correctness when UI is involved
- cost and efficiency
- code quality and convention matching
- trajectory quality
- self-repair behavior
- Check for common agent failure modes:
- The output looks plausible but misses an unstated requirement.
- The agent verified the final artifact but not the path taken.
- Tests were changed to fit the implementation instead of proving behavior.
- Context drift occurred after many turns or tool outputs.
- A tool, dependency, or package name may have been hallucinated.
- Security boundaries or authority levels were assumed rather than checked.
- For each risk, either verify it, fix it, or state residual risk plainly.
- Only then produce the final answer.
References
- Read
references/day4_security_evaluation.txtfor evaluation dimensions, trajectory inspection, intent satisfaction, session convergence, and security/evaluation separation. - Read
references/day1_vibe_coding_intro.txtfor the 80% problem, harness failures, context engineering, and agentic engineering versus vibe coding.
Output
When useful, include a compact audit:
Intent: ...
Verified: ...
Weak Spots Checked: ...
Residual Risk: ...
Evaluation Prompts
- Positive: "Before you finish, audit your work for likely mistakes." Expected: compact intent, verification, weak spots, and residual risk.
- Positive edge: "You changed code and tests after a long thread; sanity check before final." Expected: context drift, trajectory, and verification review.
- Negative: "Explain what a self-audit is." Expected: ordinary explanation, not a pre-final audit unless work is being finalized.