Postmortem

/postmortem — Retroactive Retrospective for an /execute-plan Run

Purpose: produce the same postmortem /execute-plan auto-fires, but on demand against a previously-completed run folder. Every recommendation is tagged with the postmortem taxonomy so it aggregates across runs and feeds /process-tune.

This skill is the read-only counterpart of execute-plan's postmortem step. The implementation, schemas, taxonomy, and hard rules live in the ## Postmortem section of claude/execute-plan/SKILL.md. This skill defers to that section — do not duplicate the rules here.

When to Use

• A past /execute-plan run completed without firing the postmortem (clean PASS) and you want to learn from it anyway.
• You want to re-run the postmortem after editing the run's artefacts (e.g. you added evidence to a finding, or the source PRD was updated and you want the requirements-fit lens re-applied).
• You're analysing a run produced by an older /execute-plan that pre-dates the structured-recommendation taxonomy and want to back-fill the index.

When NOT to Use

• The run hasn't completed yet — /execute-plan's postmortem step will fire if the trigger rules match. Don't pre-empt it.
• You want to re-review code quality — use /domain-review or /review-adversarial.
• Production-incident retrospective — this skill is for workflow runs, not outages. Use a different format (e.g. five-whys) for production incidents.

Arguments

Workflow

1. Locate the run folder

If <run-folder> is supplied, use it. Otherwise read runs_root: from the repo's CLAUDE.md ## Commands schema and pick the most recent timestamped subfolder.

If neither path exists, halt:

No run folder found.
Pass an explicit path: /postmortem docs/runs/<yyyy-mm-dd-HHMMSS>/

2. Verify required inputs

The run folder must contain at minimum execution-report.json. If absent, halt — there is no structured run to analyse.

Read every artefact present:

• execution-report.md and .json (required)
• disposition-log.md (optional)
• the plan file referenced by execution-report.json#plan_file
• the source PRD/AERS, if discoverable from the plan's frontmatter (source_prd:)

3. Apply the postmortem logic

Run the postmortem analysis exactly as defined in the ## Postmortem section of claude/execute-plan/SKILL.md. That section is the single source of truth for:

• mode resolution
• markdown structure (full and lightweight)
• the two named lenses (Tool and skill usage, Requirements fit)
• the recommendation taxonomy (target and type vocabularies)
• the JSON schema for postmortem.json
• the hard rules
• the headline-recommendation rule

The only difference between the auto-fired postmortem and this skill (invoked on demand):

• Trigger source. The auto-fired postmortem runs automatically based on the run's verdict and gate state. This skill fires because the operator asked. Record the operator's --reason (or "manual: <run-folder>" when none was given) in the JSON trigger field.
• Output location. The auto-fired postmortem writes adjacent to the execution report. This skill writes into the same folder. If a previous postmortem.md/.json already exists, write to postmortem-<YYYYMMDDHHMMSS>.md/.json instead — never overwrite. The index gets one row per file.

4. Append to the index

Unless --no-index is set, append a row to docs/postmortems/index.json per the schema in the ### Cross-run aggregation subsection of claude/execute-plan/SKILL.md's ## Postmortem section. Create the file as [] if it does not exist.

5. Report

Postmortem: <path-to-postmortem.md>  (mode: <mode>)
Key recommendation: <postmortem.headline>
Index: appended to docs/postmortems/index.json (run <run-id>)

If --no-index was passed:

Postmortem: <path-to-postmortem.md>  (mode: <mode>)
Key recommendation: <postmortem.headline>
Index: skipped (--no-index)

Hard rules

All hard rules from the ## Postmortem section of claude/execute-plan/SKILL.md apply unchanged:

• not a second requirements document
• not a chat-transcript summary
• not a restatement of review findings
• not a blame document
• length cap: one page (full) / half page (lightweight)
• recommendations must use the taxonomy

Plus one additional rule specific to this skill:

• Never overwrite a prior postmortem. Use a timestamped suffix on re-runs. The index relies on every postmortem being a stable, immutable artefact — overwriting one rewrites history.

Things you must not do

• Do not invent a different schema or taxonomy. The taxonomy is centralised in the ## Postmortem section of claude/execute-plan/SKILL.md so that /process-tune can group recommendations across runs.
• Do not skip the index append unless --no-index is explicit. Silent index skips defeat the auto-improve loop this skill exists to feed.
• Do not modify run artefacts other than writing postmortem*.md / postmortem*.json and appending to the index.