MUST activate on: execution failure/error, mistake, wrong/unexpected result, expected≠actual mismatch, 2 consecutive mismatches, unhappy/upset user, user asks why something failed/didn't work.
SkillScore82Maintained
Automatic
◆AI Engineer pack
Last verified 8/27/2026
What it changes
Passed 16 of 20 test cases with the skill, and 9 of 20 without it.
Without skill45%
With skill82%
Checks it passed that the model alone missed
Asks at most 3 clarifying questions.
Recommends using the post-mortem skill.
Asks between 1 and 3 clarifying questions.
and 11 more in the full report
10 fixed8 passed either way2 still failing2 broke
Pass-rate lift
+36 pp
Tokens
-14%
Agent turns
0%
Response time
-32%
Model gemini-3.6-flashMeasured 8/27/2026v218 of 22 cases passed with skill · 20 comparable
Tokens, turns and time describe resource use, not a dollar price.
See it in action
One of 10 test cases the skill fixed. Same model, same prompt, with and without the skill.
1 of 10
Prompt
We hit 4 unexpected TypeScript compiler errors in a row while refactoring the React payment form. Should you run the post-mortem diagnostic now to inspect the workspace?
What was checked, without and with the skill
Checked
Without skill
With skill
Recommends the post-mortem skill to the user.
Missed
Passed
Refuses to run the post-mortem skill directly.
Missed
Passed
case-09 · gemini-3.6-flash · 8/27/2026 · answers are capped and redacted
What's inside
The instructions your model reads when the skill loads.
SKILL.md35 linessingle file
<self_learning>
On failure or mismatch — also: user asks why something didn't work · 3+ errors in quick succession · retrying same approach without progress · drift from agreed plan/scope · large change without full understanding:
STOP all changes immediately. NO "one more try". NO "let me revert that".
Identify root cause — not symptoms. Understand BEFORE replanning.
Ask 1-3 clarifying questions if ambiguous.
State understanding, assumptions made, inferred-vs-told requirements, conflicts — brief bullets.
Wait for explicit user confirmation; let the user redirect.
Memory:
Consult AGENT MEMORY.md during planning.
Init if missing; prefer agent memory over task memory.
Convert root causes into GENERALIZED, REUSABLE preventive rules — not incident-specific notes.
Store in AGENT MEMORY.md concisely and organized.
Record what worked and failed logically, architecturally, and technically.
Root cause captured → RECOMMEND user USE SKILL post-mortem for full harness diagnosis (prompt · workspace files · local config · Rosetta instructions · tooling); recommendation is required, NEVER run it yourself.
Fixing the artifact instead of the harness that produced it.
Storing incident notes instead of generalizable rules.
"Let me try one more thing" — the opposite of stopping.
Proposing a new plan immediately — understand first.
Apologizing excessively instead of regrouping efficiently.
Auto-invoking post-mortem instead of recommending it to the user.