< all notes

The 15 Claude Code skills I keep, and why not 100

Every installed skill puts its name and description into every session, so the model can decide when to use it. That's the cost: context, and a menu that gets harder to choose from as it grows. I could have installed three full skill packs, about 100 skills. I kept 15.

What I kept

From cursor/plugins, the pstack set:

  • technical-writing for docs, READMEs, and commit messages.
  • interrogate for adversarial review by several models at once.
  • architect for sketching types and module boundaries before code.
  • show-me-your-work for a decision log on long unattended runs.
  • recall and automate-me.
  • unslop, the writing ruleset. It's set to always apply.

From gstack-distilled:

  • g-what-to-build, with YC's forcing questions and four scope modes.
  • g-ship-discipline, g-estimate-and-scope, and g-learning-loop.

From the career-ops plugin:

  • career-research and career-evaluate, which I use to read up on a company and a role before I apply.
  • career-compare and career-tailor-resume.

Separately, superpowers runs the process: brainstorm, write a spec, plan, test first, and verify before saying done.

What I left out

career-ops also ships five more, for tracking, scanning, and sending applications. I skipped all of them. I already keep one tracker, and I send my own applications. With two trackers, the model has to guess which one is real, and sometimes it guesses wrong.

The test

Before installing a skill, I ask one question: is there a task I do every week that this changes? If the honest answer is "it might be handy," it stays out. An unused skill isn't neutral. It's one more item the model reads and rules out on every turn.

Repo skills are different

Project repos carry their own. ORRN has 23 in .agents/skills/, mostly framework guides for Expo, Hono, Turborepo, and shadcn. They only load inside that repo, so the cost is paid where the benefit is.

How they combine

This site's rebuild used several at once. g-what-to-build set the scope for the second round and turned down features that wouldn't change what a recruiter does. superpowers made me write a spec and a plan first. Test-driven development pinned down the barcode encoder before any UI existed. unslop kept the copy readable. The write-up has the details.

More notes