MOC: Best AI Skills
MOC: Best AI Skills
Map of Content for the agent skills and conventions worth adopting — the ones Rufus actually keeps in ~/.claude/skills or wires into AGENTS.md. One per job: design, writing voice, slides, changelog process.
Star counts are snapshots (star_count / star_timestamp in each ref's frontmatter), not live.
The skills
Design — impeccable
Paul Bakaus's design language and agent skill that strips the "slop" from AI-generated interfaces (generic beige, purple gradients, ghost cards, side-tab borders). Hooks run each UI edit through 59 deterministic slop detectors in real time and feed findings back to the agent; the same detector can gate a PR in CI. Per-harness builds add rules for a model's known tells. Reads a project's DESIGN.md so the agent knows what a color is for.
Writing — humanizer + soundlikeme
Two jobs, two skills:
- humanizer (Siqi Chen / @blader) removes AI tells. Grounded in Wikipedia's community-maintained "Signs of AI writing" — 35 patterns across content, grammar, style, chatbot residue, and hedging. Drafts, critiques against the patterns and the original claims, rewrites. Refuses to invent the details a de-slopped rewrite tends to smuggle in.
- soundlikeme (Rufus Pollock) does the harder, less-solved half: sounding like a specific person, built from a handful of your own samples and checked with a blind-judged eval harness. Treats tell-removal as table stakes.
Removing every tell can still leave prose that sounds like nobody — and over-editing has its own detectable style. Use humanizer as the floor; use soundlikeme when the output needs to be recognisably you.
Slides — frontend-slides
Zara Zhang's skill for beautiful web presentations without knowing CSS. "Show, don't tell": it generates visual style previews and you pick, instead of describing aesthetics in words. 12 curated presets, custom animations and hover states, auto-fits any screen. Converts existing PPTX files and preserves their images and brand assets. Output is a single HTML file with inline CSS/JS — "a single HTML file will work in 10 years".
Changelog process — changelog-convention
Rufus Pollock's life-itself/changelog — a convention, not a tool, for the gap where a good changelog skill still doesn't exist. One markdown block pasted verbatim into a repo's AGENTS.md; a longer CONVENTION.md fetched only when an entry is actually being drafted. Agents draft an entry at session-checkpoint time, matched to a three-tier weight (skip / one plain sentence / full feature writeup) so the log stays readable. Entries are one file per changelog/YYYY-MM-DD-slug.md with date / title / promote frontmatter; committing to the project's own repo is the automatic floor, anything beyond that is a manual promote. Deliberately a pasted convention rather than an auto-loaded skill, so it costs no context per session.
Selection criteria
- One skill per job. Design, writing voice, slides — the shortlist, not a directory.
- Externally anchored where possible. Humanizer works from a public Wikipedia page; soundlikeme ships an eval harness. Taste that can be audited beats taste you have to trust.
- Zero / low dependencies. Markdown skills and single-file HTML output age well.
- Actually installed. These are in use, not bookmarked.
Related
- ui.sh — design agent skills from the Tailwind / Refactoring UI people.
- whizz-ai-design-prompts · claude-design · design-md — adjacent design tooling.
- voice-dna · llms-erase-linguistic-diversity — why voice matching matters.
- company-skills-library · openskills · skills-managers — managing skills at scale.
- moc-loop-engineering — skills as loop building blocks.
This MOC grows as skills earn a place on the shortlist.