Claude Code Skills

An opinionated toolkit of Agent Skills for fighting AI slop in copy, design, and code.

Each folder is a self-contained skill: a SKILL.md with instructions, plus reference files and scripts where they earn their place. The files are plain Markdown and Python, so they work in Claude Code, Codex, Cursor, and any agent that reads a SKILL.md.

Skills

Skill What it does Origin License
stop-slop Five drafting rules that remove predictable AI writing patterns: filler phrases, formulaic structures, metronomic rhythm, hand-holding, and pull-quote sentences. Gaurav Tiwari MIT
slop-detector Scores a draft or page against a four-tier severity model. Exact production narration fails outright, rhetorical scaffolds warn when repeated, lexical tells become review prompts, and document-level patterns are diagnosed. Ships a scanner, a caption linter, and regression tests. Gaurav Tiwari MIT
red-pen Strict line editor built on Orwell's "Politics and the English Language" and Butterick's Practical Typography. Returns marked-up issues with a reason per change. Rich Tabor See NOTICE.md
clarify Rewrites UX copy: button labels, error messages, helper text, empty states. Says what happened, what to do next, and cuts the rest. Adapted from impeccable Apache-2.0
distill Strips a design to its essence. The default mode of a model doing UI is to add; this skill subtracts. Adapted from impeccable Apache-2.0
quieter Tones down loud designs without flattening them: lower saturation, looser spacing, lighter weight. Adapted from impeccable Apache-2.0
polish Final pass before a UI ships: spacing, alignment, radii, hover and focus states, the small bugs that expose a rushed build. Adapted from impeccable Apache-2.0
harden Makes interfaces survive real use: long text, RTL, missing data, slow networks, error states, keyboard and screen-reader access. Adapted from impeccable Apache-2.0

Install

With the installer

git clone https://github.com/wpgaurav/claude-code-skills.git
cd claude-code-skills
./install.sh claude                      # every skill into ~/.claude/skills/
./install.sh claude stop-slop slop-detector
./install.sh codex                       # ~/.codex/skills/
./install.sh cursor                      # ~/.cursor/skills/

Set SKILLS_INSTALL_DIR to install somewhere else. The installer copies folders; it never touches skills it was not asked to install.

With symlinks

git clone https://github.com/wpgaurav/claude-code-skills ~/claude-code-skills
for s in stop-slop slop-detector red-pen clarify distill quieter polish harden; do
  ln -s ~/claude-code-skills/$s ~/.claude/skills/$s
done

Symlinks track the repository, so git pull updates every skill.

Other agents

Point the agent at the folder's SKILL.md. For project-level use, add a line to the nearest agent instruction file:

Before delivering prose, follow ./claude-code-skills/stop-slop/SKILL.md.
Before publishing, score the draft with ./claude-code-skills/slop-detector/scripts/scan.py.

Restart Claude Code after installing. Invoke a skill with its slash command (/stop-slop, /slop-detector, /clarify) or by asking for the job ("clean up this prose", "score this for AI tells", "review this error message").

Score a file

python3 slop-detector/scripts/scan.py draft.md

hard_fail (1)
  - production_commentary: 'captured with Playwright'  [block 14, figcaption]

warn_if_repeated (1)
  - repeated_scaffolds/mirrored_contrast: "It's not just faster. It's cheaper."  +2
      3 occurrences, threshold 2

contextual_review (4)
  - broad_lexical_tells: 'robust'  [block 9, body_prose]
  ...

HARD FAIL: 1 blocking construction(s). Rewrite before publishing; the score below cannot rescue them.
Advisory score from automatic findings: 2 (clean)
Ceiling if every contextual prompt is confirmed: 6 (needs editing). Review the 4 prompt(s) in context.

Matches inside quotation marks, blockquotes, and code are downgraded to review prompts, because a sentence that discusses "leverage" is not a sentence that leverages. --json gives machine-readable output, --fail-on needs_editing makes the exit code stricter for CI, and slop-detector/scripts/check-assets.py checks only captions and alt text.

Why these exist

A cold draft from any model is a smooth, confident, context-free paragraph. The first design is a gradient mesh and three pricing cards. The first commit is happy-path only. Each skill here is a checklist that runs before that output reaches you.

The rationale and worked examples are in Stop Slop Skill: 5 Rules I Use to Remove AI Slop. The design-side companion, design-slop, lives in its own repository.

Development

./scripts/validate.sh                       # frontmatter, links, private-path gate, tests
python3 slop-detector/scripts/test-severity.py

The GitHub workflow runs the same validation on every push and pull request.

Contributing

Pull requests welcome. The bar is the one the skills enforce: short, opinionated, immediately actionable. A new detector pattern needs a must-fail and a must-pass case in test-severity.py. See CONTRIBUTING.md.

License and attribution

stop-slop, slop-detector, and the repository scaffolding are MIT (LICENSE). The five impeccable-derived skills are Apache-2.0 (LICENSE-APACHE-2.0). red-pen is Rich Tabor's work, redistributed with attribution and without a license grant. Details in NOTICE.md.