failure-autopsy

Failure Autopsy

Wrong skill? Still debugging -> systematic-debugging. Stuck -> think-twice. Review feedback -> receiving-code-review.

Announce at start: "I am using the failure-autopsy skill to analyze what went wrong."

When to Use

• After an approach turned out wrong or ineffective
• After a metric was reported incorrectly
• After a ceiling or limitation was misdiagnosed
• After a regression was introduced

Example

# Example: 5-Why for misdiagnosed routing ceiling
echo "What: Routing accuracy stuck at 82%"
echo "Why1: 3 queries return wrong skill"
echo "Why2: TF-IDF scores too close"
echo "Why3: Intent patterns too generic"
echo "Why4: No collision detection in patterns"
echo "Why5: Pattern specificity never tested"
echo "Fix: Add specific patterns before generic ones"

The 5-Why Protocol

Step 1: State the Failure

What happened: [factual description] Expected: [what should have happened] Impact: [time wasted, wrong output, user intervention]

Step 2: 5-Why Chain

Level	Why?	Answer
1	Why did [failure] happen?	[direct cause]
2	Why [direct cause]?	[deeper cause]
3	Why [deeper cause]?	[systemic issue]
4	Why [systemic issue]?	[missing check]
5	Why [missing check]?	[root gap]

Step 3: Pattern Match

Pattern	Signal	Match?
False ceiling	"Impossible" proved wrong	?
Incomplete measurement	Metric on data subset	?
Confirmation bias	Evidence sought to confirm	?
Single-method validation	One way to check	?

Step 4: Preventive Action

Immediate fix + Process update + Skill update

---

Anti-Patterns

Anti-Pattern	Detection	Correction
Stops at surface	"API was slow"	Push: why was it slow? What could agent do?
No preventive action	Step 4 empty	Not done until process update exists
Same failure twice	Pattern repeats	Previous preventive action failed — escalate
Blames environment	"Network issue"	Reframe: agent-controllable factors

Failure Modes

Failure	Detection	Recovery
Stops at Why 2	<4 levels	Push deeper
No preventive action	Step 4 empty	Not done until process update exists
Same failure recurs	Pattern 2+ times	Escalate: preventive did not work
Blames externals	"API was slow"	Reframe: what could AGENT do?

Companion Skills

• think-twice: Breaking failure patterns
• systematic-debugging: Active debugging (before autopsy)
• quantitative-decision-gate: Preventing bad decisions
• measurement-integrity: Preventing bad measurements
• evolution-loop: Feeding autopsies into improvement