Claude Skills: What Each One Costs, and How Many You Can Run

How does skill loading work?

Three levels, each loaded at a different time.

Level 1, metadata. The name and description from your SKILL.md frontmatter. Loaded at startup, included in the system prompt, and present on every message. Anthropic's published figure is roughly 100 tokens per skill.

Level 2, instructions. The body of SKILL.md. Loads only when the skill triggers. Target is under 5,000 tokens.

Level 3, resources. Bundled files, templates, scripts. Nothing until accessed. Scripts are the good case: when Claude runs one, the source never enters context, only the output does.

Anthropic calls this progressive disclosure, and it works. The point of this post is what it does not defer.

Takeaway: Level 1 is a permanent cost. Levels 2 and 3 are pay-per-use.

What did the measurement show?

Twelve installed skills, measured against a live directory:

• 1,479 tokens always loaded, on every message, before you type anything

• 34,165 tokens deferred, sitting on disk until triggered

• Ratio: 23 to 1 in favour of deferral

So progressive disclosure is doing real work. For every token spent, twenty-three are saved.

The mean cost per skill came out at 123 tokens, with a range from 26 to 239. Anthropic publishes roughly 100, so the measurement and the documentation agree closely.

Takeaway: budget about 123 tokens for every skill you keep installed.

How many skills can you run?

Holding the measured mean constant:

  10 skills ->   1,230 tokens always loaded  (0.6% of a 200k window)
  25 skills ->   3,075 tokens always loaded  (1.5%)
  50 skills ->   6,150 tokens always loaded  (3.1%)
 100 skills ->  12,325 tokens always loaded  (6.2%)
 200 skills ->  24,650 tokens always loaded  (12.3%)

Three thresholds worth remembering:

Under 25: ignore it. You are spending 1.5% of your window. Install whatever looks useful.

25 to 50: prune by use. Still small. Build the habit before it costs you.

Past 100: this is real. Six percent of the window gone before the conversation starts, and it does not come back.

You will see the claim that two hundred skills cost nothing. Read strictly it is defensible, because two hundred bodies stay on disk. Read the way people actually read it, it is wrong by roughly 24,650 tokens.

On money specifically: skill descriptions sit in the system prompt, which is identical turn after turn, which is exactly what prompt caching handles. If you pay per token you are paying cache-read rates on nearly all of it. The financial cost is small. The window occupancy is the real one, and caching does not help with it.

Takeaway: the ceiling is context, not cost.

Which skills are worth the space?

The short version. Costs are measured where the skill was installed locally.

The four document skills: docx (210), xlsx (239), pptx (186), pdf (110). Together 745 tokens, and they are the difference between a real Word file and Markdown pasted into something that looks wrong.

skill-creator (83): builds other skills, and tells you when a description is too vague to fire. Cheap, and it compounds.

schedule (56): turns any prompt into a recurring task.

A brand-voice skill you write yourself: not on any marketplace, because the point is that it is yours. Largest single quality lift available, and it costs whatever you spend on the description.

Situational: mcp-builder for standing up MCP servers, webapp-testing for Playwright-driven UI tests, artifacts-builder for React and Tailwind artifacts. Install when the situation arrives, not before.

The full list with reasoning for each, including the two I removed, is in the Medium version.

Takeaway: four document skills plus skill-creator plus one custom skill covers most people at under 900 tokens.

Why is the most expensive skill worth it?

Counterintuitive, and it changed how I write descriptions.

The most expensive skill measured was xlsx at 239 tokens. It is Anthropic's own production skill. The cheapest was a setup helper at 26 tokens.

The expensive one earns the difference. It names concrete file extensions. It includes an example of how someone might mention a spreadsheet casually, without using obvious trigger words. And it lists the cases where the skill should not fire, naming what should handle them instead.

Negative triggers are the part almost nobody writes, and they are what stops skills in the same domain colliding. A description without them has no way to say "not me."

Takeaway: optimise description length by how often the skill fires, not by how tidy it looks. Verbose is correct on hot paths.

What are the frontmatter limits?

Per the documentation, as of August 2026:

name: max 64 characters, lowercase letters, numbers and hyphens only, no XML tags, and it cannot contain the reserved words "anthropic" or "claude".

description: non-empty, max 1,024 characters, no XML tags, and it must state both what the skill does and when Claude should use it.

One mechanical trap: YAML treats a colon followed by a space as a key separator, so description: Use this for: reports and memos fails to parse. Quote the whole value. The symptom is quiet — the skill still exists on disk, it just loses the metadata that would let Claude discover it.

Takeaway: 1,024 characters caps your exposure at roughly 256 tokens per skill, about double the measured mean.

Do skills work everywhere?

No, and this is the claim most often repeated wrongly.

Custom skills do not sync across surfaces. A skill uploaded to claude.ai is not available through the API. Claude Code skills are filesystem-based and separate from both. Three surfaces, three copies, and the drift is yours to manage.

On the document skills in Claude Code. This one is worth stating carefully, because a first pass at it got it wrong. The platform documentation says the pre-built document skills are not available in Claude Code. The skills repository README shows you can install the open-source versions as a plugin. Both are accurate: not available as managed built-ins, installable from the repo.

Security. Anthropic's documentation warns that a malicious skill can direct Claude to invoke tools in ways unrelated to its stated purpose, and that skills fetching from external URLs are the sharp edge, since fetched content can carry instructions. Read what you install.

Takeaway: decide which surface a skill is for before you write it.

The honest read

Four limits on the numbers above.

The tokenizer is a proxy. tiktoken failed in the sandbox (it could not reach its BPE ranks file through a restricted proxy), so this counts characters at four per token. Since tiktoken is OpenAI's tokenizer and not Anthropic's, it was an approximation either way, and the fallback has the advantage of being checkable against Anthropic's published figure. Treat everything as ±15%.

One library of twelve. Four Anthropic document skills, eight custom. Your mix will differ.

The extrapolation is arithmetic. Nobody measured a 200-skill install. The table multiplies one library's mean, which is exactly as reliable as that assumption.

Selection accuracy was not tested. There is a plausible argument that a hundred similar descriptions make it harder for Claude to choose correctly. It appears in a lot of posts as though it were established. It was not measured here and is not claimed.