Descriptions now say when each skill should fire, which words you would actually type, and when it should stay quiet. Overlaps with Claude Code's built-ins are settled in the descriptions themselves, so the choice happens at trigger time.
- ponytail vs /simplify, review-swarm vs /code-review and /security-review: tiebreaks stated in the descriptions and documented in the README.
- review-swarm stops at roughly 40 files or 2,000 changed lines, splits by area, and says which slice it read.
- architect audit findings cite both sides: the file:line in code and the doc that should describe it.
- ask-the-council weighs the cost of being wrong before seating a panel.
- up-to-date handles a branch with no upstream instead of reporting a failed check as in sync.
- prompt-generator adds a ground-truth section so generated prompts pin verified paths and versions.
- README gains a companion-skills section for ux-designer and humanizer.