Skillit: keep your Claude Code skills lean and cheap

I run mycaminoguide.com, an AI guide for people walking the Camino de Santiago. It's mine. I pay the bills, which means every token it spends is my money, not some company's cloud budget. So when I started building Claude Code skills to ship features faster, I got a nasty surprise the first time I actually looked at where the tokens were going.

One skill I'd written was fine. It worked. It also reloaded its entire bloated self into context on every single run, quietly charging me rent I didn't know I was paying.

Why skills bloat, and why that's expensive

A skill isn't free just because it sits in a folder. Its description loads in every session whether you use the skill or not. The body loads every time the skill triggers. And if you stuffed all your rules, templates, and edge cases straight into SKILL.md, all of that rides along too, every time.

Write a few of those and the cost compounds. I wrote up the manual fixes a while back in best practices to make Claude Code skills token-efficient, and the numbers genuinely surprised me. Moving detail out of one skill body into reference files cut it from about 4,760 tokens to 1,930, a 60% drop, without deleting a single thing. On a real Postman plugin task, the same discipline made the run 27% cheaper and 46% faster.

That post is the recipe. The problem is that following a recipe by hand, every time, on every skill, is exactly the kind of thing you stop doing when you're moving at founder speed.

What Skillit actually does

So I built Skillit, a Claude Code plugin that bakes those best practices in so I don't have to remember them. It's three pieces:

• skill-builder, a guided tutor that walks you from a half-formed idea to a working, validated skill.
• /skillit:skill-new, which scaffolds a new skill with the lean structure already in place.
• /skillit:skill-validate, which scores an existing skill and tells you exactly what's bloating it.

It's built on top of Anthropic's own skill-creator, not instead of it. Skill-creator cares whether your skill produces good output. Skillit cares whether your skill is structured well and cheap to run. Different jobs. I use both.

Three tiers and progressive disclosure

The core trick is progressive disclosure, the same one from the best-practices post, now automatic. Every skill Skillit scaffolds has two layers:

my-skill/
  SKILL.md        # frontmatter + the core workflow, loaded when the skill runs
  references/     # the deep detail, loaded only when Claude actually needs it

SKILL.md stays small: a tight description and the workflow steps. The heavy stuff, the templates, the long rules, the examples, goes in references/ and sits there costing nothing until the moment it's needed. You keep all your documentation. You just stop paying for it on every run.

Scoring the sizzle

The part I lean on most is /skillit:skill-validate. It grades a skill against six things: description quality, body size and focus, progressive disclosure, tool scoping, frontmatter validity, and prompting craft. You get a number back and a reason for it. Twelve out of twelve is Sizzling 🍳.

One detail I'm kind of proud of: the scoring runs in a separate agent, so grading your skill doesn't dump a rubric into your main context and cost you the very tokens you're trying to save. It would be a bad look to bloat your session while telling you to slim down.

Tool scoping is the criterion that catches the most people. In the best-practices post I cut one agent from 111 allowed tools to 11. Wildcards feel convenient and hand every component the entire toolbox. Skillit flags it.

Install it

It's two commands inside Claude Code:

/plugin marketplace add quintonwall/skillit
/plugin install skillit@skillit

It's free and open source, so if you want to read the fine print first, it's all on GitHub.

Skills aren't just something you build. They're something you learn.

Skillit isn't a black box that spits out a file and sends you on your way. It's a tutor. As it builds, it explains the choices: why this lives in the frontmatter and that gets deferred, why a tighter description saves you tokens on every single run, where a skill is doing too much.

The first skill you make with it will be leaner than anything you'd have written alone. The tenth, you'll write mostly yourself, because by then the instincts are yours. That's the real return: not one optimized skill, but a sharper sense for how to build with Claude that compounds across everything you ship next.

And for me, that return has to pay rent. mycaminoguide.com runs on my dime, so every token I don't waste on a bloated skill is a token I get to spend on something a pilgrim actually feels. Less spent on the plumbing, more left for the thing you're actually trying to build.