Skills for Claude Code, extracted from systems that run in production rather than designed in the abstract.
Each skill is a self-contained directory: its own SKILL.md for Claude, its own README.md
for you, and its own installer if it needs to set up a project repo.
| Skill | What it does |
|---|---|
| copy-council | Multi-viewpoint review of website and landing page copy. Five seats that cannot see each other's work read the page against a Core Idea settled with you first, and one of them gates the verdict rather than the average, because clean sentences do not rescue a page that argues badly. |
| eng-doc-review | Adversarial review of an engineering plan before anyone builds it. Five independent angles read the doc against the actual codebase, then consolidate into a severity-ranked scorecard whose verdict is the lowest angle rather than the average. |
| humanize | Detection and removal of AI writing tropes, run as an edit pass rather than as a rule list in the writer's context. Three parallel editors and a second-pass scanner cover 40 patterns across diction, structure, and tone, then a humanize pass rewrites whatever survives into colleague-to-colleague voice. The trope list never reaches the writer, because telling a model not to use a pattern is a reliable way to get the pattern. |
| loop-and-harness | A queue-driven development loop. Claude works through a queue of goal contracts one at a time, promoting, building, verifying, merging, and closing out each before the next, with mechanical gates rather than good intentions. |
| pm-doc | The full lifecycle of a product document: a persona probe that establishes who the work has to serve, a premise challenge, a section-by-section interview into a draft, and a five-lens readiness review scored for design, engineering, and agent handoff. |
| storytelling | Narrative structure for documents that have to persuade: presentations, RFCs, strategy memos, product pitches. Five frameworks, ten named anti-patterns, and an audit that scores the argument rather than the prose. |
git clone https://github.com/superstrong/agent-skills.git
cd agent-skills
./install.shThis copies every skill into ~/.claude/skills and does nothing else. It never touches a
project repo, writes a cron entry, or changes any config.
./install.sh loop-and-harness # just one
./install.sh --helpSet CLAUDE_SKILLS_DIR to install somewhere other than ~/.claude/skills.
Some skills also set up individual project repos. Those ship their own install.sh, which
the root installer names but never runs for you, since they do more than copy files. See each
skill's README.
Issues and pull requests are welcome.
To add a skill, give it a top-level directory containing:
| File | Required | Purpose |
|---|---|---|
SKILL.md |
Yes | What Claude loads. Frontmatter needs a name and a description specific enough to route on, since the description is what decides when the skill triggers. |
README.md |
Yes | Orientation for a human: what it does, why, how to install it, what it requires. |
install.sh |
If it sets up a repo | Per-repo setup. Accept --help. Do not assume the skill was installed system-wide first. |
references/, assets/ |
As needed | Progressive-disclosure docs, and files the skill copies into a target repo. |
Keep the root layer generic. Anything specific to one skill belongs inside that skill's directory, including its installation instructions.
MIT. See LICENSE.