Fewer skills, on purpose

Before this, I ran an OpenClaw-style skill catalog. On the order of 12,000 skills. In about three months it drew roughly 138K views. Traffic said keep going. Using it said stop.

The problem wasn't missing pages. It was the opposite. Ten near-clones for every real job. Fresh SKILL.md files every week. Nobody — including me — could honestly answer which one to install. Competition was pure volume: whoever indexed more won the SEO race and lost the trust race.

Then James McLaughlin bought a $5 coffee (opens in a new tab). Five dollars. After months of warehouse inventory. That wasn't funding. It was a clean read of the room: keep a personal map of what I trust, or become an unpaid full-time janitor for other people's half-finished prompts.

I chose the map.

This isn't only a personal burnout story. Research in 2026 keeps hitting the same wall. Four findings that matter here:

• Curated skills help on average — SkillsBench (opens in a new tab) (+16.2 pp pass rate across 7,308 trajectories), but the spread is wild by domain and 16 of 84 tasks get worse. Self-generated skills give almost nothing on average (−1.3 pp). Focused packs of two or three modules beat long documentation-as-skill.
• Piling on docs isn't free — When Skills Don't Help (opens in a new tab) finds procedural docs barely move a tool-grounded CTF setup (~+8.9 pp, not significant). Sometimes the environment already supplies the feedback skills are supposed to provide.
• Authored skills still need human repair — SkillAxe (opens in a new tab) finds LLM-authored skills often sound fine and still fail; a small fixed library can beat a giant unread pile. Not All Skills Help (opens in a new tab) measures causal heterogeneity: some help, many do nothing, some hurt.
• Agents invent skill names — Skills That Don't Exist (opens in a new tab) shows fake names ~36% of the time; names that repeat across models become easy attack targets.¹

People on X live this without reading abstracts. One builder shipped a Raycast browser because skills get fun until you have too many with cryptic names and no clue which to pick (opens in a new tab). Security work keeps counting corpses in the marketplaces — roughly 1 in 4 skills scanned with vulns, ~1 in 20 malicious (opens in a new tab); scanners over tens of thousands of OpenClaw-style skills surface injection and leaked secrets (ClawSecCheck notes on 27k skills (opens in a new tab)). Others put the poison rate even higher: ~20% of an OpenClaw marketplace poisoned, 1,184 skills stealing from agents that installed them (opens in a new tab). Catalogs that celebrate a million skills sit next to people calling the same layer a malware-filled mess (opens in a new tab).²

I got tired of being on the “more index entries” side of that fight.

The core insight: judgment beats volume.

What stays is tools I open in a normal week. Skills that change outcomes, not just token counts. Plugins I'd reinstall after a wipe. References that still hold when you reread them. Right now that's 44 entries. If something leaves the setup, it leaves the site.

Want every skill that ever had a markdown file? Plenty of other URLs for that. This one is for sharing what I actually use — and what I'm willing to put my name on.