# Debrief

> Use at the very end of a delivery run, after knowledge maintenance and the recorded outcome, to run the propose-only learning pass: reading the run's journal, identifying process-level friction (gates that never bite, questions the human answered repeatedly, rules that misfired, rework loops), and proposing Intelligence-Layer improvements — every proposal gated on explicit human approval, with a changelog entry on apply. This is the process_retrospective action — one Agent-tool dispatch with subagent_type debrief — it studies how the run went, never what it built, and it edits nothing itself.

- Skill: `bhangeef16/debrief` (Agent Skill)
- Install (CLI): `npx skillmds@latest add bhangeef16/debrief`
- Raw SKILL.md: https://api.skillmd.com/api/skills/bhangeef16/debrief/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: AI & ML
- Author: BhangeeF16 (https://skillmd.com/u/bhangeef16)
- Updated: 2026-09-17
- Page: https://skillmd.com/skills/bhangeef16/debrief

---


# Debrief

Close the run by learning from it. Product learnings were captured by the knowledge-maintenance actions before this; this action's subject is the process itself.

## Dispatch

One synchronous `Agent`-tool call, `subagent_type: debrief`. The dispatch prompt names the run's journal location (`.kaizen/runs/<run-id>/journal.md`, per the delivery-lead's own journaling duty).

## Required inputs

- The completed run's journal — one line per dispatch, checkpoint decision, and human escalation, appended live by the delivery-lead during the run.

## Required outputs (all mandatory)

- A proposal list — per proposal: the observed pattern, its journal citation, the target file, the proposed minimal change, and the expected effect on future runs — OR an explicit no-proposals record naming the journal read and the patterns checked.

## Rules — the propose-only boundary

- No proposal is applied without an explicit human decision on it; an unanswered proposal is not an approved one.
- Applied proposals get a changelog entry recording what changed and why; rejected proposals are recorded so they are not re-raised every run.
- A missing journal is a stop condition — say so plainly; never reconstruct one from memory or transcript fragments.
- A proposal that would change a mandatory human-approval checkpoint or a security-related rule is flagged high-impact and requires explicit discussion, not just a yes/no.

## Result

Report the proposal list (or no-proposals record), the result (pass / pass_with_risks / blocked / failure), and blocking questions as an explicit list (empty list stated explicitly).