Portable agent skills for Claude Code and compatible harnesses, installable with one command.
# Everything
npx skills add graseeel/skills
# A single skill
npx skills add graseeel/skills --skill brutal-review
# Global (user-level) install
npx skills add graseeel/skills -gManual fallback — clone the repository and copy the skill folder into ~/.claude/skills/ (or your harness's skills directory):
git clone https://github.com/graseeel/skills
cp -r skills/brutal-review ~/.claude/skills/
cp -r skills/dumb-comments ~/.claude/skills/| Skill | What it does |
|---|---|
brutal-review |
A brutal, round-based adversarial code review that ends only on a complete clean round. |
dumb-comments |
A comment-only skill that makes every comment pay rent; code is never touched. |
Adversarial code review: fresh reviewers with zero context attack the diff every round, a skeptic kills every finding without a concrete failure scenario, and the loop ends only on a full round with zero confirmed findings (6-round cap).
- When to use: a change must survive a merciless review, or you need a verdict on a pull request you cannot modify.
- Modes:
brutal(fix loop) andverdict(one-pass, findings only). - Requirements: git for the diff scope; independent subagents recommended, with a documented sequential fallback.
Comment-only cleanup: every comment pays rent or loses its line. Narration, ceremony, and AI voice are deleted or rewritten in the file's own voice; behavior change is zero by construction, verified mechanically on the produced diff.
- When to use: comments in a diff or file set look AI-written or useless.
- Modes:
apply(default) andreport(assessment only, nothing edited). - Requirements: none — single-agent; comment detection uses the file's own language syntax.
Skills are portable SKILL.md files: no producer-machine dependencies, no external parsers, no language-specific tooling. They load in any harness that reads SKILL.md skills.