Claude Code skills for SEO: what to install and what not to automate

Claude Code skills are a good fit for the SEO work that has a right answer, and a bad fit for the SEO work that has a defensible answer. That line runs straight through the middle of the discipline, and which side a task falls on predicts whether automating it helps or quietly hurts.

The useful version of this post is not a list of skills to install. It is the test for deciding, because the list changes monthly and the test does not.

The test: can the output be checked without you?

A skill is worth building or buying when its output can be verified by something other than your own judgement at the moment you read it.

JSON-LD passes: you generate it, then you run the URL through Google's Rich Results test and the answer is binary. A robots.txt audit passes: the user agents are documented by each operator, so a skill can check your file against the current crawler list rather than against what someone wrote in a blog post in 2023. An internal link check passes: the link resolves or it does not.

Keyword selection fails the test, and fails it badly. A skill will produce a keyword map, complete with priorities and rationale, from no demand data at all. It will look exactly like a good keyword map. There is nothing to check it against, and you will not find out it was wrong for two months.

That asymmetry is the whole argument. Automating verifiable work compounds. Automating judgement work produces confident output that feels like progress and defers the error.

What belongs in a skill

Four categories, in the order I would build them:

Schema generation and validation. The structured data docs define what each type needs, and the failure modes are mechanical: missing required fields, wrong nesting, an entity referenced by an id that nothing defines. A skill that generates markup and then validates the rendered page is doing real work. One that only generates has automated the easy half and left you the half that catches errors.

Crawler and access audits. Whether AI and search bots can read you is a yes-or-no question, it changes when operators rename user agents, and almost nobody re-checks it. Perfect skill material. The five-minute version by hand is the same checklist.

Link and structure integrity. Internal links that resolve, headings that nest, canonical tags that point somewhere real, metadata inside its length limits. Boring, checkable, and the category where manual review reliably fails because the work is tedious rather than hard.

Page-level content structure. Whether every H2 is followed by a self-contained answer, whether each section carries a specific number. Structural, checkable, and distinct from whether the content is any good.

What to keep by hand

Keyword selection. Covered above. The model has no demand data, and the confident output is the danger.

Deciding which page owns which query. This is a decision about your whole site, made against competitors, with consequences that take months to show up. A skill sees one page.

Competitive positioning. What to say that nobody else is saying requires knowing what everybody else is saying, which requires reading them.

The pattern: anything where being wrong is invisible for eight weeks should stay slow and human.

The overlap problem, with numbers

SEO is the worst domain for skill collision, and I can put a number on it from my own install. On September 21, 2026 I counted 383 installed skills. Twenty-six had names beginning with seo and another twenty with geo, plus thirty-one beginning with blog that overlap on content work.

Those are not forty-six capabilities. They are forty-six near-neighbours competing to match the same requests, and the skills documentation makes clear that selection runs on descriptions. When several descriptions plausibly fit, the model picks one, and which one is close to arbitrary.

The practical effect is worse than it sounds. You install a good technical-audit skill, it works, you later install a good on-page skill, and the technical one quietly stops winning requests it used to win. Nothing errored. I have written up the mechanism separately, because it applies to every domain and hits SEO hardest.

The fix is not fewer skills so much as descriptions written against each other, with each naming what it handles and what it leaves to its neighbour.

The thing skills genuinely cannot do

Worth being blunt about, because it is the gap most SEO tooling pitches talk around.

Skills replace workflow and checking. They do not replace data. Search volume, keyword difficulty and backlink indexes come from crawlers and clickstream panels that cost real money to operate, and no amount of skill design conjures them. If your tool subscription is mostly for the data, a skill is not a substitute and anyone telling you otherwise is selling.

What a skill does replace is the part of the subscription you were paying for out of habit: the audit checklists, the schema builders, the crawl reports you could run yourself and did not because it was tedious. That is a real saving, and it is smaller than the pitch.

For the audit-and-checklist half done properly, AI SEO Command Suite is the pack I built around this split, and Structured Data Pro Pack is the narrower one if schema is the only piece you need. If you are unsure whether you need a pack at all, start with the free route, which is the right answer more often than I would like.

Where to start

1. Audit crawler access by hand once, so you know what the skill should be checking.
2. Add one schema skill, and make it validate rather than only generate.
3. Add a link and structure checker, and run it in CI rather than on request.
4. Leave keyword work alone until you have demand data to check it against.

Step four is the one that will be hardest to keep to, and the only one on this list where ignoring me costs you two months instead of an afternoon.

Related: the six packs I actually use, and where to find skills worth installing.