Nine times out of ten the description is too abstract. Add the three essentials: what it does (list concrete capabilities), when to use it, and the words users might actually say (cover casual phrasing). Then verify by making a matching request.

Run /memory first: the rule may live in a subdirectory file that has not lazy-loaded yet. Then check for two rules contradicting each other (Claude may pick either). For hard constraints that still wobble, rewrite them as a PreToolUse hook.

Performance drops as context fills up. Cut unrelated tasks apart with /clear; on long tasks run /compact focus on the essentials at ~70% full; persist conclusions to CLAUDE.md or let auto memory (MEMORY.md) take over — never rely on chat history.

Resident cost is tiny (~100 tokens of metadata each), but overlapping descriptions blur matching. Keep only the best one per niche, and let each skill do exactly one thing.

Hard facts and rules → CLAUDE.md; multi-step "how-to" → Skill (loaded on demand); actions that must be enforced → Hook; connecting external data and systems → MCP, then use a Skill to teach Claude how to use those tools well.