PrimeSkills — a curated marketplace for Claude skills and AI agents. Every listing goes through my own review before it ships: does it work outside a demo? Does it solve a real problem?
After 130+ listings reviewed, these four are the ones I still reach for every week. Not because they're the most impressive — because they replaced something genuinely painful.
primeskills
1. Defuddle — stop feeding Claude cluttered HTML
The problem: When I use Claude to research documentation or read articles, WebFetch returns everything: navigation bars, cookie banners, sidebar widgets, ad wrappers. Claude processes all of it. The actual article might be 20% of the tokens consumed.
What the skill does: Defuddle is a CLI that extracts clean, readable markdown from any web page — removing everything except the content.
defuddle parse https://some-docs-site.com/guide --md
The skill teaches Claude to prefer Defuddle over WebFetch for standard web pages. The output is clean markdown. No nav, no footer, no clutter.
Real improvement: Docs research sessions with Claude got noticeably faster and cheaper. For pages with heavy surrounding boilerplate (most enterprise docs), the actual content is often 30-40% of the raw HTML — Defuddle brings that back to 90%+.
→ Defuddle on PrimeSkills
2. LLM Wiki — give Claude a memory that compounds
The problem: I do a lot of ongoing research — reading papers, tracking how specific AI techniques evolve, keeping up with tooling. Every new Claude session, I'd start from scratch. Re-explaining context. Re-reading the same sources. The knowledge didn't accumulate.
What the skill does: Based on Andrej Karpathy's LLM Wiki pattern, this skill builds and maintains a persistent knowledge base as interlinked markdown files.
The structure is three layers:
-
raw/ — immutable source material (articles, papers, transcripts)
-
entities/ and concepts/ — agent-maintained wiki pages
-
SCHEMA.md + index.md + log.md — structure and navigation
When you give it a new source, it reads the index first, checks what already exists, then creates or updates pages with cross-references. Contradictions get flagged. Synthesis reflects everything ingested. The wiki is just a directory of markdown files — open it in Obsidian or VS Code.
Real improvement: This is the difference between a research assistant that forgets everything and one that builds on what it knows. A week in, the cross-references start to get useful. A month in, querying the wiki is faster than re-reading the papers.
→ LLM Wiki on PrimeSkills
3. Design Taste (Anti-Slop Frontend) — Claude-built pages that don't look AI-generated
The problem: Every landing page Claude builds defaults to the same aesthetic: Inter font, purple-blue gradient, three identical feature cards in a row. You can spot it from a mile away. It's not that the code is wrong — it's that the design choices are completely generic.
What the skill does: This skill intervenes before Claude touches any code. It reads the brief and infers a design direction: page type, audience, vibe, appropriate design system. It sets three explicit dials (Design Variance, Motion Intensity, Visual Density) and picks the right stack and aesthetic family.
Specific guardrails it enforces:
- Avoids Inter as the default font — picks Geist, Satoshi, Cabinet Grotesk based on brief
- Bans the AI-purple gradient unless the brief explicitly asks for it
- Prevents "three equal feature cards" layout
- Maps each brief to a real design system (shadcn/ui, Radix, Fluent, Carbon) when applicable
- Has a pre-flight checklist of ~40 things to verify before shipping
Real improvement: The first time I used it to build a landing page, the output chose Geist Display, an asymmetric hero, and a neutral zinc palette with an amber accent. Not perfect, but not immediately recognizable as AI-generated. That's the bar.
→ Design Taste on PrimeSkills
4. Android Perfetto Analysis — trace debugging in an hour instead of a day
The problem: Perfetto is the right tool for Android performance debugging — but it requires knowing the right SQL queries against a specific schema, understanding which metrics to run first, and knowing which domain hints to follow for CPU vs. memory vs. jank issues. The learning curve is real.
What the skill does: A Google-authored skill that implements a structured investigation protocol:
- Initialize a chain-of-evidence scratchpad in the trace directory
- Review domain hints (CPU, Memory, I/O, IPC, Graphics, Power) to understand what's possible
- Read the SQL reference and follow its execution protocol — no guessing at schemas
- Follow a depth-first investigation loop: hypothesize → collect data → drill down
The skill knows not to stop at the first anomaly. Complex performance issues rarely have a single cause — it keeps digging until it's confirmed there are no other major bottlenecks.
Real improvement: Perfetto analysis that used to require someone with trace expertise now has a systematic protocol. You still need to provide the trace and know roughly what symptom you're investigating — but the "where do I even start?" friction is gone.
→ Android Perfetto Analysis on PrimeSkills
What makes these work
All four share something: they're not trying to do everything. Defuddle does one thing (clean web content extraction). LLM Wiki has strict rules about what goes where. The frontend skill has a 40-point pre-flight checklist. The Perfetto skill has a mandatory investigation loop.
The pattern I keep seeing in skills that actually survive production: tight scope + strict protocol + real documentation. Skills that try to be general-purpose tend to be mediocre at everything.
The catalog at primeskills.store has 130+ listings now — all reviewed against that bar. If you're building with Claude and want to skip the trial-and-error phase, it's worth a look.
What's the most tedious part of your current Claude workflow? I'm always looking for gaps in the catalog — drop it in the comments.