SuperAPI AI Directory

retro

Agents & models

View repository

Updated 10-15 14:47

About this skill

Retro Skill

YOU MUST EXECUTE THIS WORKFLOW. Do not just describe it.

Quick-capture a learning to the knowledge flywheel. For comprehensive retrospectives with backlog processing, activation, and retirement, use /post-mortem.

Loop position

Lightest capture surface for move 7 of the operating loop. Single-observation quick-capture into .agents/learnings/. Use when an insight is too small to warrant a full /post-mortem and too potentially-useful to leave in the handoff. The promotion ratchet still applies after capture — a retro entry is a candidate learning, not yet a pattern or rule.

Quick Mode

Given /retro --quick "insight text" or /retro "insight text":

  1. Generate a slug from the content: first meaningful words, lowercase, hyphens, max 50 chars.
  2. Resolve the active bead with timeout_run 1 bd current 2>/dev/null || echo "".
  3. Write directly to .agents/learnings/YYYY-MM-DD-quick-<slug>.md:
---
type: learning
source: retro-quick
source_bead: <active bead id when available>
source_phase: validate
date: YYYY-MM-DD
maturity: provisional
utility: 0.5
---

# Learning: <Short Title>

**Category**: <auto-classify: debugging|architecture|process|testing|security>
**Confidence**: medium

## What We Learned

<user's insight text>

## Source

Quick capture via `/retro --quick`

If no bead is active, omit source_bead intentionally and still set source_phase: validate.

  1. Confirm:
Learned: <one-line summary>
Saved to: .agents/learnings/YYYY-MM-DD-quick-<slug>.md

For comprehensive knowledge extraction, use `/post-mortem`.

Done. Return immediately after confirmation.

Examples

User says: /retro --quick "macOS cp alias prompts on overwrite — use /bin/cp to bypass"

What happens:

  1. Agent generates slug: macos-cp-alias-overwrite
  2. Agent resolves the active bead with timeout_run 1 bd current 2>/dev/null || echo ""
  3. Agent writes learning to .agents/learnings/2026-03-03-quick-macos-cp-alias-overwrite.md with provenance fields like:
---
type: learning
source: retro-quick
source_bead: bd-123
source_phase: validate
date: 2026-03-03
maturity: provisional
utility: 0.5
---
  1. Agent confirms: Learned: macOS cp alias prompts — use /bin/cp

Troubleshooting

Problem Cause Solution
Learning too generic Surface-level capture Be specific: "auth tokens expire after 1h" not "learned about auth"
Duplicate learnings Same insight captured twice Check existing learnings with grep before writing
Need full retrospective Quick capture isn't enough Use /post-mortem for comprehensive extraction + processing

Reference Documents

  • references/retro.feature — Executable spec: single-observation capture to .agents/learnings/, candidate-under-ratchet, lightest move-7 surface (soc-qk4b)