Skip to content
Dustin's AI Lab|AgentCrew AgentCrew
Go back

This post is archived: it may be out of date and is kept for reference.

Second Harness Diet: Global Skills From 58 Down to 40

A follow-up to the harness diet series: six months later, a bigger cleanup that cut my global skills from 58 down to 40, start to finish.

Last time I wrote about the harness diet, I’d trimmed my personal Claude Code setup by −36%. Six months on, I ran a second round, bigger this time: a big consolidation of my global skills, from 58 down to 40.

The approach was to send in two Sonnet subagents to take inventory first, one sweeping the 58 at the user level and one sweeping the 66 at the project level, then I made every call myself once they were done.

Twelve went straight into the archive at ~/.claude/skills/_archived/: meeting-debrief, slack, teach-impeccable, frontend-slides, job-tailor, browser-routing, third-party-quotas, obsidian-markdown, obsidian-bases, obsidian-cli, polite-biz, humanizer-zh. Six business-book skills got merged into a single unified-routing biz-books. A few overlapping ones got folded together outright: darwin into skill-creator, explore-unknowns into first-principles, humanizer’s Chinese and English versions into one, polite’s two modes into one. artist-sentiment moved over to the crawler project. Two got downgraded into lighter forms: browser-routing into a resident rules/common/, third-party-quotas into a reference inside memory.

At the end I added spoken-language triggers to 9 skills, ran the whole lot through full YAML-format and ≤200-character validation and passed, and fixed up all the cross-references too: CLAUDE.md, tool-patterns, error-recording, plus pdf, obsidian, social-card, threads-voice, launch-kit.

That’s it.


Free Prep: Get Your First AI Result

Step-by-step setup for Claude Code with zero coding required.

Join Free Prep

Ready for AI to finish a real task for you? Explore L1 Not sure where to start? Book a free 15-min chat

Share this post on:
Previous Post
Two Traps Running Claude Code on Local Ollama: Truncated Context, and A3B Buckling Under Heavy Verification
Next Post
Picking a Brand Look as a Non-Designer: AI Made Prototypes Cheap, So Taste Became the Hard Part