Skip to content
Back to skills

Retrospective

ASecurity

Cycle-end retrospective converting recurring judgment failures into durable enforcement. Reviews the cycle's friction — repeated user corrections, guard trips, RCA themes, integrate failures — and turns each recurring one into a landmines.md entry, then proposes graduation candidates up the enforcement funnel (landmine → advisory hook → hard gate). Pairs with standup: standup says what shipped and what's next; retrospective says what kept going wrong and how to stop it structurally. Writes on...

  • 14 stars
  • 0 votes
  • 0 copies
  • 1 view
  • Added October 4, 2026
ai-agentsgo

Works with

  • cli

Security analysis

A100/100

Scanned October 4, 2026

npx -y skills add friedbotstudio/baseline --skill retrospective --agent claude-code

Installs into .claude/skills of the current project.

Are you the author of Retrospective?

Add the live security badge to your README. It updates with every re-scan.

Security grade badge for Retrospective
[![Security: A — Skills Directory](https://www.skillsdirectory.com/api/skills/friedbotstudio-retrospective/badge)](https://www.skillsdirectory.com/skills/friedbotstudio-retrospective)

More formats (shields.io, HTML) on the badges page. Keep it an A: scan every change in CI with Pro.

Download with Pro
SKILL.md
---
name: retrospective
owner: baseline
description: Cycle-end retrospective converting recurring judgment failures into durable enforcement. Reviews the cycle's friction — repeated user corrections, guard trips, RCA themes, integrate failures — and turns each recurring one into a landmines.md entry, then proposes graduation candidates up the enforcement funnel (landmine → advisory hook → hard gate). Pairs with standup: standup says what shipped and what's next; retrospective says what kept going wrong and how to stop it structurally. Writes only memory entries and a report — graduation proposals are advisory and require the normal seed.md amendment path to become enforcement. Invoke at the end of a release cycle, an epic, or any stretch of work with repeated friction.
disable-model-invocation: true
---

# retrospective — from recurring friction to structural enforcement

The baseline's enforcement funnel has three rungs: a **landmine** (a remembered failure pattern surfaced by advisory memory hooks), an **advisory hook** (a PreToolUse hook that surfaces the landmine inline before the risky call), and a **hard gate** (a blocking hook or consent gate). Friction that recurs should climb this funnel instead of being re-litigated every session. This skill is the climb's entry point.

**This skill is a generator, not a workflow phase.** It enters no Track Guard ordering and blocks nothing. It writes ONLY memory entries (via the canonical entry shape) and an inline report. It never edits hooks, never touches `seed.md`/`CLAUDE.md`, never creates enforcement — those are amendment-path changes the user drives.

# Inputs

Gather in main context (Article II — no subagent judgment):

- The cycle window: since the last release tag, the epic's first commit, or a user-named range.
- The `landmines` and `decisions` categories under `.claude/memory/` — what is already recorded. Read them shape-agnostically: a sharded store keeps one fact per file under `<category>/`, a flat one keeps `## key` blocks in `<category>.md`.
- Friction evidence inside the window: guard-trip messages in `.claude/state/logs/` (hook log lines), `docs/rca/*.md` postmortems, integrate-failure yields in `.claude/state/harness/*.log`, and user corrections you can cite verbatim from the session or `_resume.md`.

# Method

## 1. Collect candidate frictions

Sweep the inputs for events that happened **more than once** in the window, or once with high cost (a broken commit, a failed release, a consent near-miss). One-off annoyances are not retro material.

## 2. Classify each recurring friction

- **Already a landmine?** Re-verify the entry (Article IX.2); refresh `last-touched` and strengthen the entry body with the new occurrence.
- **New pattern?** Draft a `landmines.md` entry in the canonical shape (stable key, occurrence citations, `verified-at`, verbatim quote when the source is a user correction).
- **Not memory-shaped** (a spec gap, a missing test)? Route it to `backlog.md` as an entry instead — the retro records it; a future workflow fixes it.

## 3. Propose graduations

For each landmine with ≥ 2 recorded occurrences ACROSS cycles, assess the next rung:

| Current rung | Graduation | What it takes |
|---|---|---|
| landmine entry | advisory hook surfaces it inline (e.g. `process_lifecycle_guard` pattern) | seed.md §4.1 amendment + hook change — user-driven |
| advisory hook | hard block / consent gate | same amendment path, stronger justification |

A proposal names: the landmine key, occurrence count, the proposed rung, the hook/gate shape, and the cost of NOT graduating. It is **advisory** — put it in the report, not in any enforcement file.

## 4. Report

```
# Retrospective — <window>

## Recurring frictions (N)
- <key> — <occurrences, cost, where recorded>

## Memory writes
- <landmines.md/backlog.md key> — new | strengthened | re-verified

## Graduation candidates (advisory)
- <key> — landmine → advisory hook: <one-line shape> — evidence: <counts>

## Not retro material (dropped)
- <item> — <why>
```

# Constraints

- **Memory writes only** — `landmines.md` / `backlog.md` / `decisions.md` in the canonical entry shape, respecting `size-cap: 500` and verbatim-provenance rules (Article IX). No other file writes.
- **No enforcement changes.** Proposing is this skill; amending is the user's seed.md §4.1 path.
- **Evidence or it didn't happen.** Every friction cites its occurrences (log line, RCA path, commit, or verbatim correction). No vibes-based landmines.
- **Read-only outside memory.** No state files, no workflow.json, no consent paths.

Attribution

Is this your skill, or is something wrong with this listing? Request removal or report an issue. Author removals are honored within 72 hours.

Comments

Loading comments…

Newsletter

Get the weekly list of safe new skills, and the ones we pulled.