Building a Skill Library

Introduction

Knox has 60+ skills in ~/.claude/skills/. That number is a feature, not a bug — it means there is an expert workflow encoded for nearly every task he runs regularly. But 60 skills only compounds if the library is organized, triggers are precise, and curation removes skills that no longer earn their place.

A skill library that grows without curation accumulates dead weight: trigger collisions, outdated verification gates, skills that were written for a workflow that no longer exists. The curator lifecycle is what separates a library that compounds from one that degrades.

Organization by Domain

Domain organization is not just aesthetic — it's functional. Skills in the same domain share a trigger namespace, which means trigger names can be designed to be distinct within a domain without worrying about cross-domain collision. /ship (Build & Ship domain) and /retro (Context & Memory domain) can coexist without confusion because their trigger phrases live in entirely different semantic spaces.

Knox's five domains cover the entire workflow: Build & Ship for delivery, Quality & Audit for code health, Content & Writing for the content operation, Context & Memory for session continuity, and Academy & Tracks for platform work.

Within each domain, skills are ordered by invocation frequency. The high-frequency skills get the most curation attention; the low-frequency ones get reviewed during the monthly audit.

Trigger Reliability Analysis

The fire rate analysis reveals two failure modes:

False fires — the skill fires when the user didn't intend it. Usually caused by short, common trigger words like "learn" or "research." Fix by adding discriminating words: "extract lessons," "what did we learn" are more specific than "learn."

Dead skills — the skill never fires because its triggers are too specific or don't match how users naturally phrase the request. Fix by looking at what the user actually typed and adding that phrase as a trigger.

Knox monitors this informally — when a skill doesn't fire on a natural request, he adds the natural phrase to the TRIGGERS section. The skill file is the living record of how the workflow is actually invoked.

The Curator Lifecycle

The curator lifecycle is the monthly discipline that keeps the library healthy. The key metrics to check:

Fire count per skill (last 30 days): A skill that fires 0 times is either dead (retire) or has bad triggers (fix). A skill that fires 20 times per month gets hardened — its verification gate must be airtight.

False fire rate: How often did the wrong skill fire? This surfaces trigger collisions and over-broad trigger phrases.

Gate failure rate: How often did a skill claim done before the verification gate actually passed? This surfaces lazy gates that need tightening.

The curator review takes 30-45 minutes monthly and pays for itself in reduced friction — every trigger collision resolved is a wrong-skill invocation prevented.

Naming Conventions That Scale

Skill file names follow three rules in Knox's library:

1. Match the primary trigger — the file name is the slash command form. If the skill is invoked as /audit-swarm, the file is audit-swarm.md.

2. Domain prefixes prevent collision — skills in the same area use a distinguishing prefix when needed. The academy skills prefix with academy-, the agent skills prefix with agent-.

3. Verb-first naming — audit-swarm.md, ship.md, learn.md, retro.md — verbs signal what the skill does. Noun-first names like security.md or build.md tend to generate trigger conflicts.

Scaling from 5 to 60 Skills

The transition from a few skills to a large library happens gradually, and it's where most operators accumulate technical debt. The pattern Knox follows:

0-10 skills: Add freely. Build confidence with the format. Don't worry about organization.

10-25 skills: Trigger collisions start appearing. Introduce domain groupings. Add a naming convention.

25-50 skills: Monthly curation becomes necessary. Fire rates reveal dead skills. Archive the first cohort.

50+ skills: The library is an asset that requires maintenance like any codebase. Monthly curation is non-negotiable. Add a skills CLAUDE.md that documents the domains and naming conventions for future reference.

The library becomes more valuable with scale, but only if the curation keeps pace. A 60-skill library that isn't curated is harder to use than a 10-skill library that is.

What's Next

The next lesson is the capstone — assembling settings.json, hooks, skills, and commands into a coherent personal harness, and running the audit lab that verifies every layer is healthy and earning its place.