Your Claude Code skills cost 7,155 tokens on every request

I installed a few plugins. Then a few more. Then I measured what they cost.

7,155 tokens. On every single request. Whether I use them or not.

That is 64 skills across 21 sources, and I had no idea, because nothing in the interface tells you. The /context command shows what your conversation is using. It does not itemise what your tooling is using before you type a character.

Here is the mechanic, the measurement, and three things I found that I did not expect.

Why a skill you never use still costs you

A skill is a folder with a SKILL.md. The body — the actual instructions — loads only when the skill fires. That part is pay-per-use and fine.

The frontmatter is not. Claude has to decide which skill to invoke, and it decides by reading every skill's name, description and when_to_use. So those get injected into the system prompt of every request, for every installed skill, forever.

Ten plugins at eight skills each is eighty descriptions riding along on every turn. Nobody notices, because each one is small and the total is never shown.

  64 skills across 21 sources
  ~7,155 tokens added to every request

  BY SOURCE
  --------------------------------------------------------------
  personal                      3,439 tok   48.1%  ████████████
  plugin-dev@claude-plugins-of    766 tok   10.7%  ███
  pentest@claude-pentest          426 tok    6.0%  ██
  mcp-server-dev@claude-plugin    347 tok    4.9%  █
  project-artifact@claude-plug    231 tok    3.2%  █
  injection-probe@talhayme        211 tok    2.9%  █

Half of my cost comes from one source. I would not have guessed that, and guessing is the only tool most people have here.

Three things the numbers showed

1. I had the same skill installed twice

  ⚠ OVERLAPPING SKILLS
    [accuracy, analysis, benchmark] skill-creator (personal),
                                    skill-creator (skill-creator@official)

A personal copy and the official plugin's copy, both loaded, both describing the same work. Not just wasted tokens — when two skills compete for the same request, Claude picks one semi-arbitrarily. The other is pure overhead, and which one wins is not something you control.

2. Descriptions get silently truncated

Claude Code cuts name + description + when_to_use at 1,536 characters. Past that, the text is dropped.

This is worse than it sounds. The author wrote instructions describing when their skill should fire; the tail of those instructions never reaches the model. The skill still loads. It just gets chosen on incomplete information, and nothing warns anyone — not the author, not you.

3. The heaviest skills are all description, no restraint

   252 tok  docs
   241 tok  pptx
   240 tok  google-workspace
   239 tok  xlsx
   237 tok  computer-use

Roughly 240 tokens each, on every turn, forever. A description's job is to help Claude choose — that needs a sentence or two. Everything past that belongs in the skill body, which loads only when the skill actually fires.

What about MCP? (this part is out of date everywhere)

If you search for this problem you will find posts about MCP servers burning 55,000 or 67,000 tokens of tool definitions before you type anything. Those numbers were real.

They are also substantially out of date, and most of the posts repeating them do not say so. Claude Code shipped MCP Tool Search in v2.1.7: when tool descriptions would exceed roughly 10% of the context window, definitions are tagged for deferred loading and replaced with a search tool. The model searches, loads only what matches, then calls. Reported reduction is around 85%.

So the MCP half of this problem has a built-in fix.

Skills do not get that treatment. Their descriptions are how Claude decides what to invoke, so they cannot be deferred behind a search without breaking selection. The skill tax is still paid in full, every turn, and it is the part nobody is measuring.

Measure your own

I wrote a plugin for this. It reads the on-disk layout — personal skills, every installed plugin, project-local skills — and reports the total, attributes it per source, and flags the three problems above.

claude plugin marketplace add talhayme/claude-plugins
claude plugin install context-audit@talhayme

Then ask in plain language — "how much context are my skills costing me?" — or run /context-audit directly.

Offline, read-only, no dependencies. Token counts are estimated at ~4 characters per token: good enough to rank and budget, not exact.

For CI, there is a budget flag, so a repository's .claude/skills cannot grow unnoticed:

python3 scripts/audit.py --budget 4000   # exit 1 when over

What I actually did about it

Uninstalled the duplicate skill-creator. Went through the personal skills — the 48% — and removed what I had installed to try once and never used again.

I did not go further, and this is the part worth saying out loud: 7,000 tokens is not an emergency. On a 200k context window it is around 3.5%. The reason to measure is not that the number is catastrophic; it is that the number was invisible, and invisible costs only grow. Mine went from 6,630 to 7,155 while I was writing this post, because I installed two more of my own plugins.

Which is the honest footnote here: my three plugins cost 525 tokens between them. A tool that measures overhead is still overhead. If context-audit is not earning its 132 tokens on your machine, uninstall it after you have your number — it will tell you the same thing next time you reinstall it.

The plugin is context-audit, MIT, part of a small marketplace of tools built on one idea: AI tooling should be measurable.