prune-skillslisted

# Pruning the skill files Skill files grow in one direction. Every edit adds the reasoning that motivated it, in the house style, which rewards concrete detail — and nothing in the workflow ever removes any of it. `--tools` deletes claims that have gone *false*; it is silent on prose that is verbose and true. The cost is not disk. A skill file is loaded in full the moment it is invoked, so every retained paragraph is paid again on every invocation, forever. An agent file is worse: it is loaded on every dispatch to that agent, in a context the session never sees and cannot correct. ## The `--prune` shorthand `--prune` invokes this immediately, with no confirmation. It grants **COMMIT, not push** — which this skill never exercises itself, because it writes nothing: the grant is what `--tools` inherits when the cuts are dispatched to it, so the cuts land in a commit rather than only in the working tree. Where a flag counts, and the authorization it confers, live in `shorthand-flags.sh` and [`~/.claude/README.md`](../../README.md) — one copy, not restated per skill. ## Scope The files are **`record.tooling.sources`** in the project's `.claude/workflow.json` — the same set `--tools` owns, so the two skills can never disagree about what a tooling file is. Without a manifest, fall back to `.claude/skills/*/SKILL.md` and `.claude/agents/*.md` and say in one line that you did. Both scopes count where a project has both. A repository whose sources include globally-installed ski