The moment you learn Claude supports skills, everything starts looking like a skill. A write-PR-descriptions skill. An explain-this-regex skill. A draft-release-notes skill. Most of those ideas are good, but not all of them earn the overhead.

A skill is a bet: you're betting the cost of writing, maintaining, and debugging a SKILL.md is less than the cost of re-prompting the same way every time. That bet pays off for repeatable specialist work. It doesn't pay off for one-shot tasks or for things the model already does well with a plain prompt.

The three tests a skill should pass

Repeat-use. You'll ask the same kind of thing more than a handful of times. Writing a skill for a once-a-year task is overhead without a payoff.

Specialist shape. The task has a non-obvious recipe — "here's what a safe migration looks like," "here's how we format release notes," "here's the checklist for an accessibility audit." Something you'd reteach every new hire.

Stable output shape. You want consistent output across invocations. Without a skill, two invocations of "draft release notes" will produce subtly different formats. A skill nails the template.

When it's not a skill

If the task is one-shot ("fix this typo"), the answer is just prompt it. If it must run every time some event happens, regardless of what the model thinks, the answer is a hook (see the primer). If the task is long-running and would bloat the main context, the answer is a subagent — a separate agent with its own context window that hands back a short report.

Eight live scenarios

Below are eight realistic asks. For each one, think about whether you'd write a skill, add a hook, spawn a subagent, or just prompt. Then reveal to see where I'd land and why.

Anatomy of a skill worth keeping

A good skill is short and opinionated. Its description is the discoverability contract — Claude reads descriptions across all available skills when deciding whether to invoke one. Make that description a sentence a new reader would scan and say "yes, that's what I want."

The body is the playbook: what to do, in what order, with pointers to scripts or reference files the skill loads on demand. Keep the always-in-context part slim; push reference material into sibling files loaded only when actually used.

A test that saves skills from bloat

Once a quarter, open each of your skills and ask: "would I write this skill today, knowing what I know now?" The answer is sometimes no. A skill that made sense six months ago can become cruft — the model got better at the task, your product changed shape, the recipe drifted. Delete it. Your future turns will be cheaper and Claude will stop mis-invoking the outdated one.

Skills are tools, not trophies. The ones worth keeping are the ones you'd rewrite from scratch.