| name | evaluateskills-postmortem |
| description | Automatic post-mortem after each /evaluateskills run — compares each mutated skill to the baseline archive pre-run to detect dormant narrative regressions. Invoke automatically from Step 5 of evaluateskills, with ARCHIVE_PATH and MUTATED_SKILLS as inputs. Do not invoke manually except to re-audit a past run. |
| narrative_critical | true |
Skill: Evaluateskills Postmortem
3rd line of defense against narrative drift in skills. Compares each skill mutated during the current run to its state before the run (Step 0 archive), via read-only Haiku agents.
Why this is necessary: the dry-run (Step 3.6 of evaluateskills) compares the current original to the mutated version during the run — it does not see cumulative drift over multiple runs. This skill compares the pre-run archive to the final version after the run, from session memory. This is what detects dormant bugs like the closeday bug of 2026-04-15 (5 mutations each scraping 1 point, below the dry-run's radar).
Required Inputs
Invoked with:
ARCHIVE_PATH — absolute path of the archive created at Step 0 (e.g., {VAULT_PATH}/{CLAUDE_CODE_FOLDER}/Skills/Archives/2026-04-16_14-30/)MUTATED_SKILLS — list of skills mutated during the run: names + absolute paths to current versions
If ARCHIVE_PATH is missing or invalid: notify {USER_NAME} and skip the postmortem without blocking Step 6.
Step 1 — Parallel Audit (Haiku, read-only)
For each skill in MUTATED_SKILLS, spawn a background Haiku agent in parallel. Agents do not modify any files.
Agent prompt (model: Haiku):
You audit the changes made to the skill: [skill name]
Current file (post-run): [absolute path current]
Baseline file (pre-run): [ARCHIVE_PATH]/[skill-name].md
Begin by reading both files.
MISSION: identify differences between baseline and current, categorize each one.
Categories:
- BUG_FIX — correction of incorrect behavior
- EDGE_CASE — coverage of missing edge case scenario
- REFACTOR — restructuring without functional impact
- WORDING_NEUTRE — reformulation without loss of information
- NARRATIVE_REMOVAL — removal of instructions that condition output richness (grouping, multi-day retrospective, open questions, pivotal moments, analytical dimensions)
- LEAN_AGGRESSIF — reduction that removes functional content, not just noise
- UNCERTAIN — change with ambiguous impact on output
If frontmatter contains `narrative_critical: true`: lower detection threshold for NARRATIVE_REMOVAL — any prescriptive instruction removed is to be flagged, even if it seems redundant on the surface. These instructions exist for a reason: they condition the output.
VERDICT per change: KEEP (improvement or neutral) | RESTORE (probable regression)
RETURN FORMAT (single line if CLEAN, structured block if ALERT):
[name]: [categories detected] | VERDICT: CLEAN
[name]: NARRATIVE_REMOVAL, LEAN_AGGRESSIF | VERDICT: ALERT
→ RESTORE: [precise diff — removed lines, with sufficient context for surgical rollback]
If an agent fails or returns nothing: log [postmortem FAILED] [skill name] and continue without blocking.
Step 2 — Aggregation and Decision
Wait for all agents, then:
If all verdicts are CLEAN:
Nothing to display to {USER_NAME}. Silently append to skills-autoresearch-log.md:
[postmortem] Run [date] : N skills audited — no regression detected.
Let evaluateskills continue to its Step 6.
If ≥1 verdict ALERT:
Display to {USER_NAME} before moving to Step 6:
⚠️ Postmortem — regressions detected in this run
| Skill | Categories | Content Removed |
|-------|-----------|-----------------|
| [name] | NARRATIVE_REMOVAL | [precise diff] |
For each skill in ALERT:
- RESTORE → restore from archive: cp "[ARCHIVE_PATH]/[skill].md" "Skills/[skill].md"
- KEEP → accept the mutation despite the signal (to justify)
- INSPECT → display full diff for manual decision
Wait for {USER_NAME}'s decision skill by skill. Do not move to Step 6 until each ALERT has received a response.
