Agent skill · civitai
cleanup
Runs the two no-gate review lanes over everything this session touched, consolidates them, and APPLIES the fixes: comment-review (comment necessity + false claims) and docs-drift-review (docs the session made wrong). Session-scoped, not segment-scoped. Use partway through a long session, and before opening a PR.
What it needs
About 4k tokens when loaded.
What this skill does
Cleanup — the two lanes nothing else checks Runs comment-review and docs-drift-review over the whole session's work, consolidates them into one list, and applies the fixes — the lanes produce exact replacement text, so a run that stops at a list has done the expensive half and skipped the useful one. Agent Reviews --- --- comment-review Comments: still true? earns its place? said in the fewest words? docs-drift-review Docs the session made wrong — stale paths, done checklist items, contradictions — plus padding Both lanes are read-only and neither has an automated gate — comments aren't type-checked and docs are never executed, so typecheck, lint, prettier and every test suite pass over both. That is the whole reason this exists. These two lanes are pre-authorised. Running them needs no separate permission, and neither does re-running one that goes idle. That covers these two only; everything else about spawning is unchanged. 1. Scope to the SESSION, not to a segment This is what makes it different from civitai-review and svelte-review, which scope to a slice of work. Here the scope is everything that has happened since the session started: 🔴 A shared worktree may contain someone else's work. More than one agent can be editing the same tree, and files you never touched will show up in git status. Before spawning, separate what this session did from what it did not, and pass only the former. Reviewing another agent's half-finished edits produces findings nobody asked for and can send them backwards. If unsure whether a file is yours, say so and leave it out — a named exclusion is fine, a wrong inclusion is not. 2. Fan out Give each agent the file list and the commit range. They need different things beyond that: comment-review — the diff. It reviews comments in the change; adjacent noise is at most a one-line note, so it needs to know which lines are new. …
How to use it
Reference it in AdaL, Claude Code, Cursor or any coding agent — nothing to install:
@skills civitai/cleanup