Prune

Condenses the markdown docs and puts them back on convention, and sweeps the skill registry. Git is the real archive — nothing here is lost by being cut.

When to use this

• Periodic housekeeping; not gated on any other skill finishing.
• Skip if the docs are already within the thresholds below — the checks are cheap, the rewrites are not.

Each doc's job

File	Holds
README.md	how to run it
CLAUDE.md	invariants
docs/COMPLETE_ACTIONS.md	append-only log
docs/plans/*	NNN-short-slug.md, zero-padded, never reused — NNN is the chronological id

Content in the wrong file gets moved to the right one, not deleted.

Steps

1. Survey before proposing. Line counts, plan count, and any file that doesn't match the naming convention.

   wc -l README.md CLAUDE.md docs/COMPLETE_ACTIONS.md
   ls docs/plans/
   

2. Apply the thresholds.

   • docs/COMPLETE_ACTIONS.md over 300 lines → compress everything older than the last 3 entries to one line per entry. Never drop an entry entirely.
   • More than 10 plans in docs/plans/ → move the oldest to docs/plans/archive/, leaving open and unstarted plans where they are. The index row stays in docs/plans/README.md; archiving is not a status change.
   • Any plan not named NNN-short-slug.md → rename it and update every reference in the other markdown files.

3. Remove what's redundant or outdated — content the code now contradicts, or that a later entry supersedes.

4. Remove closed reviews — specs/*/review.md with status: closed whose plan is already implemented.

5. Sweep the skill registry (skills/README.md). This is the tooling as written — file-level checks only. For the tooling as used (which skills actually fired, where sessions stalled), that is /observer, which reads transcripts; don't duplicate it here. Read every row and flag any skill or agent that is:

   • not referenced in docs/skill-workflow.md's pipeline map and not tagged overlay or utility — an orphan;
   • last-reviewed older than 90 days — stale, needs a look;
   • over the ~120-line size smell in its SKILL.md (wc -l);
   • overlapping another core skill per the Overlap check in skills/README.md — flag the pair for consolidation.

   Present the flags. On a confirm-keep, bump that row's last-reviewed to today. Delete a skill only on explicit confirmation.

6. Deep-audit a single skill — when the run names one skills/<name>/SKILL.md rather than the whole registry, additionally check, citing SKILL.md line numbers with a severity on each finding:

   • Safety — allowed-tools / tools wider than the Procedure uses; a destructive or outward-facing action (delete, force-push, send, deploy) with no confirmation step.
   • Triggering — if the complaint is "it doesn't fire", compare its description trigger phrases against the request that failed and against sibling skills' triggers for a collision that routes elsewhere.
   • Standard — section order, constitution restatement, size smell, per the authoring standard in skills/README.md. Report findings; apply fixes only on confirmation, same as the registry sweep.

7. Present the proposed actions as a list and get confirmation before touching anything. This skill rewrites history-shaped files; it doesn't do that unasked.

What not to do

• Don't delete a COMPLETE_ACTIONS entry or a plan file — compress and archive instead.
• Don't renumber a plan, ever, including an abandoned one.
• Don't move content into CLAUDE.md that isn't an invariant; it is not a changelog.