| name | retrospective |
| description | Context engineering audit after mistakes. Use when user says "retrospective", "retro", "introspect", "context audit", "what went wrong", or asks to analyze/prevent session mistakes. Invoke IMMEDIATELY when these triggers appear - don't do ad-hoc analysis. |
| user-invocable | false |
Retrospective
YOUR ENTIRE VISIBLE OUTPUT MUST BE ONE OF THESE THREE TEMPLATES. NOTHING ELSE.
TEMPLATE A — Gaps found
**BLUF: All addressed in-session (X/Y).** ← or "X/Y addressed in-session." if partial
**Unaddressed:**
1. Add "X rule" → file.md (gap: did Y instead of Z)
**Addressed:**
1. Add "X rule" → file.md (gap: did Y instead of Z) ✅ applied
**Session:** One sentence.
**Closure: Safe to close.** ← or "NOT safe." with numbered list of unfinished items
If all gaps are addressed, omit the Unaddressed section. If none are addressed, omit the Addressed section.
TEMPLATE B — No gaps, non-trivial session
**No gaps.**
**Session:** One sentence.
**Closure: Safe to close.** ← or "NOT safe." with numbered list of unfinished items
TEMPLATE C — No gaps, trivial session
**No gaps.**
NO TEXT BEFORE THE TEMPLATE. NO ANALYSIS. NO EXPLANATION.
Scope Rule
Two sections, both mandatory:
Section 1: Context Engineering Gaps
Rules missing, rules not followed, knowledge to persist. This is the gap analysis.
Section 2: Session Closure Readiness
Can user close this session without losing work? Check for:
- Uncommitted code changes (
git status)
- Unpushed commits
- Unsubmitted QA (changes deployed but no QA message sent)
- Unfinished work promised this session (e.g., "I'll write that test" but didn't)
- Pending builds/deploys that haven't been verified
If ALL clear → append **Closure: Safe to close.** after the Session line.
If NOT clear → append **Closure: NOT safe.** Unfinished: with a numbered list.
This prevents the false "all done" signal that caused the user to almost close a session with 5 unfinished items (Mar 2026 incident).
Post-Template Behavior
Default: Auto-address immediately.
After outputting the template:
- If Unaddressed items exist → immediately proceed to fix them (edits, updates, etc.)
- After each edit, VERIFY it landed → grep/read the file for the new content. If not found, the edit failed — retry. Never report "addressed" without verification evidence.
- If an item needs user input → ask only for that item, continue with others
- Report results after all auto-addressable items are done, using this format:
"X/Y addressed (found + finalized N gap(s)). All done." — where X/Y is the final tally and N is how many gaps were discovered and fixed during the auto-address phase (gaps that weren't already addressed in-session). If no new gaps were found during auto-address, just "X/Y addressed. All done."
Exception — Pause mode:
If user explicitly says "retrospective and pause", "retro only", "just report", or similar → output template only, don't auto-address. Wait for user to say "address it" or similar.
Anti-Confabulation Rule
🚨 NEVER fabricate, misquote, or reconstruct prior outputs from memory. When referencing what you previously said (counts, quotes, claims), either:
- Cite it exactly with the message context, or
- Say "I don't have the exact text" and describe the gist
Incident (Mar 2026): In a meta-retro, Claude claimed "I said 4/4" when the actual output was "3/4 addressed." Under pressure of being challenged, fabricated a worse version of its own error to construct a coherent narrative. This is confabulation — inventing false details when uncertain. It destroys trust faster than the original mistake.
Rule: If challenged on what you previously reported, don't guess. Say what you know and what you don't.
