A Claude SEO skill is a folder with a Markdown file in it. The file tells the agent what the skill does, when to use it, and how to run the job. Claude Code, Codex, and other agents that support the convention read the same format: a short metadata block at the top, then the instructions.
The question every team hits about two weeks in is which of these to install. Install everything and you worry about filling the context window. Install nothing and you are back to copying prompts out of a notes app.
We measured our own library to answer it, and the answer splits cleanly in two. A skill costs almost nothing while it sits unused. It costs real context when it fires. The median skill in our library is 107 tokens of standby metadata and 1,851 tokens of instructions. Those two numbers should be doing different jobs in your decision.
The measurement
We parsed every SKILL.md file in the two skill roots on one machine and split each file at the frontmatter boundary.
Measure | Value |
|---|---|
Skill files parsed | 76 |
Distinct skill names | 75 |
Total metadata block | 8,231 estimated tokens |
Total instruction body | 222,952 estimated tokens |
Body to metadata ratio | 27 to 1 |
Token counts are estimated at four characters per token, which is a good approximation for English technical writing but not exact. Treat them as comparative, not literal. The ratio and the distribution are the findings; the absolute number will move with your own files.
Why the shelf cost stays small
Skill systems use progressive disclosure. The metadata block is loaded so the agent knows the skill exists and when it applies. The body is read only when the skill is invoked.
That design produces a distribution worth looking at.
Percentile | Standby cost per skill | Body cost per skill |
|---|---|---|
Smallest | 43 tokens | 428 tokens |
Median | 107 tokens | 1,851 tokens |
90th percentile | 161 tokens | 7,125 tokens |
Largest | 246 tokens | 21,702 tokens |
The whole 76 file library carries 8,231 tokens of standby metadata. That is roughly the size of a long article, and it is the price of the agent knowing that 75 skills exist before it has done anything. The largest single metadata block in the library is 246 tokens.
The max is the interesting one. The most verbose skill in the library spends 246 tokens telling the agent when to use it, which is still small enough that a shelf of skills will not sink a session. If you are worried that installing five SEO skills will crowd out your data, the measurement says the shelf is not where the crowding happens.
Where the cost actually lives
The body is where it gets real, and the spread is wide. 15 of 75 skills load under 1,000 tokens when they fire. 11 load over 5,000. The largest loads 21,702.
For an SEO program, that range is the thing to plan around, because it maps to how much work the skill does. A skill that audits one page is cheap. A skill that runs a full site crawl, compares a baseline, and writes a report has a lot of instructions because it has a lot of steps.
Invocation cost clusters at the low end and then runs long, with 11 skills above 5,000 tokens.
Here is our search and measurement set, measured the same way.
Skill | Standby | When it fires |
|---|---|---|
tool-cluster-builder | 70 tokens | 428 tokens |
seo-tools-local | 102 tokens | 638 tokens |
geo-operator | 101 tokens | 665 tokens |
dataforseo-toolkit | 238 tokens | 849 tokens |
ahrefs-2026-content-refresh | 51 tokens | 958 tokens |
search-engine-visibility-audit | 49 tokens | 1,121 tokens |
codex-link-equity-audit | 96 tokens | 1,163 tokens |
gsc | 116 tokens | 1,309 tokens |
bing-webmaster | 135 tokens | 1,685 tokens |
seo-beginner-automation | 98 tokens | 1,836 tokens |
All ten together | 1,055 tokens | 10,652 tokens |
Ten SEO skills cost about 1,000 tokens on the shelf and about 10,000 tokens when all of them fire.
Two things fall out of that table.
The standby column is nearly flat. Ten SEO skills cost 1,055 tokens to have available. That is a rounding error next to a page of search results or a crawl export, and it is the argument for installing the shelf you want rather than the shelf you think you can afford.
The invocation column is not flat. The set ranges from 428 to 1,836 tokens, a factor of four. In practice you rarely fire more than one or two in a session, so the honest planning number is not 10,652. It is a few thousand tokens on top of whatever data the session is already holding.
What the numbers mean when you choose a skill
Three selection rules that follow from the measurement.
Judge the body, not the description. The metadata block is the skill's advertisement. A skill can have a 50 token description and a 7,000 token body. Read the file before you install it, and check the length of everything below the frontmatter. If you want a quick filter, file size and line count are a decent proxy for how much the skill will pull in.
Prefer skills that read data over skills that embed it. A skill that shells out to an API and reads the response costs a fixed amount regardless of how many keywords you have. A skill that carries a lookup table or a template library inside the Markdown costs the same whether you need the table or not. This is why a 849 token DataForSEO wrapper is cheaper in practice than it looks, and why a big template-driven content skill is more expensive than its description suggests.
Budget the session, not the shelf. The number that matters is total tokens in an active session, and skills are one input among several. A session that has a crawl export pasted in and three skills loaded is a context problem. A session with twenty skills available and one loaded is not.
Where a skill is the right choice, and where it is not
Skills and MCP servers solve different halves of the same problem, and using the wrong one for the job is why teams end up with neither.
A skill is instructions and knows nothing. It is the right shape for a repeatable judgement, such as how to triage a ranking drop, how to structure an audit output, or what evidence a link profile review needs. Its cost is context, and it is paid per session.
An MCP server is a live connection. It is the right shape for data the agent cannot otherwise reach, such as Search Console or a backlink index. Its cost is configuration, credentials, and permissions, and it is paid once. If you want to know what those connections actually expose before wiring one up, our comparison of four SEO MCP servers is in this probe.
The pair that works: one connection per data source, one skill per recurring judgement. In our own workflow the Search Console connection is a server and the daily monitoring routine is a skill, because the judgement is the part that repeats.
What to check before you install one
A short verification pass, because the format is easy to write and easy to write badly.
- Read the frontmatter and confirm the description names a specific trigger. A description that says "helps with SEO" will load at the wrong moments.
- Count the lines below the frontmatter. Under 100 lines is a focused skill. Over 500 lines, check whether the detail is reference material that should be a separate file the skill reads on demand.
- Search the file for credentials and hardcoded paths. A skill should read configuration from the environment, not carry a token.
- Check whether it writes to your site. A skill that only reads is safe to try on a live property. A skill that edits files should have a review step before anything changes.
- Run it once on a property you do not care about and read the output. Most skills are judged on their output format, not their instructions.
FAQ
Do installed skills slow down every session? Only by their metadata blocks. In our library the median is 107 tokens and the total across 76 files is 8,231 tokens. The instructions load when the skill is used, not before.
How many SEO skills should I keep installed? In our own set, ten search and measurement skills cost about 1,000 tokens of standby metadata. The limit is rarely context. The real limit is your own ability to remember what each one is for.
Is a large skill file a bad sign? Not automatically. Reference-heavy skills are legitimately long. But length should be justified by what the skill does. If a 500 line skill runs a single API call with a lot of explanation, that is a documentation problem, not a capability.
Can I write my own Claude SEO skill? Yes, and it is the highest-return option once you have a routine that repeats. Start by running the job manually twice, write down the steps that did not change, and put those in the body. Keep the frontmatter under 150 tokens and describe the trigger precisely.
Do these numbers apply to Claude Code specifically? The format and the progressive disclosure model are shared across agents that support SKILL.md, including Claude Code and Codex. The measurement is from our own library, so expect your distribution to differ with your file sizes.
Auspia view: install the shelf you want, then audit what fires. The standby cost of a skill library is small enough to stop worrying about, and the invocation cost is where a badly written skill actually hurts. If you have not measured your own set, counting lines below the frontmatter in each file takes about ten minutes and tells you where the weight is. The detailed skill-versus-server split for GEO work is covered in our GEO skill notes.
Author: Alice Monroe, AI SEO Tools Analyst Covering 150+ Tools at Auspia. Alice writes about AI SEO tooling, skill and server architecture, and what each layer of an agent stack actually costs.
