Stub: retro skill (steering-hygiene lens only)

Matt tweeted the shape on Aug 15-16 (compound engineering's promise
without ceding control to the model): mine session friction, propose
CODING_STANDARDS.md additions AND deletions, land as a PR rather than
writing directly. Name, destination file, and the "must delete too"
requirement all came from that thread.

Three more lenses (tool economy from session JSONL, navigation-pointer
drafting, multi-session transcript fan-out) are named in the SKILL.md's
"Not yet built" section but not specified closely enough to build yet.

In-progress bucket: no docs page, no plugin/README (top-level)
registration, no ask-matt update needed per CLAUDE.md's bucket rules.

.changeset/retro-stub.md

@@ -0,0 +1,5 @@
+---
+"mattpocock-skills": patch
+---
+
+Add `retro` to `in-progress` (stub): mines session friction into `CODING_STANDARDS.md` proposals, runs the deletion test against existing standards so it prunes as well as adds, and lands everything as a PR — never a direct write, since the model proposing its own constraints needs a human gate, not a formality. v1 covers the steering-hygiene lens only; tool-economy, navigation-audit, and multi-session lenses are named but not built.

skills/in-progress/README.md

@@ -14,3 +14,4 @@ npx skills@latest add mattpocock/skills --skill=<name>
 - **[writing-shape](./writing-shape/SKILL.md)** — Take a markdown file of raw material and shape it into an article paragraph by paragraph, arguing format choices at each step.
 - **[claude-handoff](./claude-handoff/SKILL.md)** — Hand the current conversation off to a fresh background agent that picks up the work immediately, seeded with a handoff summary via `claude --bg`. User-invoked.
 - **[setup-ts-deep-modules](./setup-ts-deep-modules/SKILL.md)** — Wire dependency-cruiser into a TypeScript repo so each package is a deep module — implementation hidden in subfolders, reachable only through its entry-point files, tests exercising it through those. User-invoked.
+- **[retro](./retro/SKILL.md)** — Mine a session for friction, propose it as a `CODING_STANDARDS.md` diff (additions and deletions), land it as a PR — never a direct write. Stub: only the steering-hygiene lens is built. User-invoked.

skills/in-progress/retro/SKILL.md

@@ -0,0 +1,40 @@
+---
+name: retro
+description: Retrospective on how a session went — mine friction into CODING_STANDARDS.md, prune what no longer applies, and land it as a PR.
+disable-model-invocation: true
+argument-hint: "Nothing, to retro this session — or point it at a past transcript"
+---
+
+# Retro
+
+Stub: v1 of the **steering hygiene** lens only, the one specified in enough detail to build. Three more lenses were sketched while designing this skill and aren't built yet — see the bottom of this file.
+
+This is the compound-engineering promise — the process improves over time — without compound engineering's failure mode: models are bad at improving their own behavior unsupervised. So `retro` never edits a steering file directly. It proposes, on a branch, as a PR. A human reads the diff before any standard changes.
+
+## Friction
+
+The raw material of a retro pass: a place in this session where you corrected the agent, re-explained something already true, or the agent visibly floundered — retried a tool call, misread a file, took a wrong turn before self-correcting. Mine the conversation for these first. Don't invent friction that didn't happen.
+
+## Routing a piece of friction
+
+Not every finding is a `CODING_STANDARDS.md` candidate. Picking the wrong destination just adds noise to the wrong document:
+
+- **CODING_STANDARDS.md** — a rule about how code should be written or reviewed, enforceable by `/code-review` at review time. The default destination, and the only one this skill should propose without asking.
+- **AGENTS.md / CLAUDE.md** — a navigation miss, not a behavior miss: the agent burned turns finding something a pointer would have handed it directly. Flag it, don't write it — these files stay hand-curated, a few highways, not a log.
+- **CONTEXT.md** — a domain-language gap. Flag it for a `domain-modeling` pass rather than writing it inline; that skill's active build/sharpen discipline is a different shape of work than a retro proposal.
+
+If a piece of friction fits none of the three, drop it. A retro that force-fits everything into `CODING_STANDARDS.md` is the CLAUDE.md-bloat failure mode this skill exists to avoid.
+
+## Process
+
+1. Mine friction from the current conversation (see above).
+2. Route each finding per the section above. Only the `CODING_STANDARDS.md` candidates get written up as a diff; anything else becomes a short list for the human to act on separately, not part of the PR.
+3. Read the repo's `CODING_STANDARDS.md` in full, not just the new candidates, and run the deletion test against every existing rule — call the Skill tool with "codebase-design" for its exact wording. A rule earns removal when deleting it changes nothing: the behavior it once corrected doesn't recur, or a newer rule already subsumes it. Propose deletions alongside additions in the same diff. A retro that only adds becomes the mess a standards file turns into when nothing is ever cut.
+4. Draft the diff. Branch, commit, open the PR, and hand the human a summary of what's added, what's cut, and why for each. Never write directly to `CODING_STANDARDS.md` — the whole point of routing everything through review is that the model is the one proposing its own constraints, so the human is the gate, not a formality.
+5. List any navigation or domain-language findings that survived step 2 after the PR link. They are not part of the diff.
+
+## Not yet built
+
+- **Tool economy** — read the session transcript for expensive or repeated tool calls and propose cheaper tool setups. No reference vocabulary exists yet for this anywhere in the repo.
+- **Navigation audit** — go further than flagging (see Routing above) and actually draft the `AGENTS.md`/`CLAUDE.md` highway pointers, rather than leaving them for the human.
+- **Multi-session mode** — fan out sub-agents across several past transcripts in the same project before mining friction, so one session's idiosyncrasy doesn't get promoted to a standing rule. The shape: parallel Explore sub-agents, one per transcript, findings reconciled before routing.

skills/in-progress/retro/agents/openai.yaml

@@ -0,0 +1,5 @@
+interface:
+  display_name: "Retro"
+  short_description: "Turn session friction into standards, safely"
+policy:
+  allow_implicit_invocation: false