Claude Code SEO Skills: What One Real Test Found

You will learn: what a skill is; how it differs from a plugin and a subagent; a safe way to try a public SEO skill; what its checklist can and cannot establish; and how to check later whether an AI answer cites the edited URL.

What is a Claude Code SEO skill?

In Claude Code, a skill is a folder with a SKILL.md instruction file. The file describes when the workflow applies and what Claude should do. You can call it by name, or Claude may load it when the task matches its description. Anthropic's skills documentation explains where personal and project skills live and how they load.

A skill is not a source of search-volume data by itself. It can read a file you provide, inspect a page in your repository, or tell Claude to use an approved tool. The quality of the output depends on the instructions, the evidence available and your review. A social video showing a polished report does not reveal whether the underlying checks ran, whether external data was licensed, or whether the author accepted the edits.

A plugin can package skills alongside agents, hooks and other components. A subagent is a separate worker Claude can assign a task. A skill may ask Claude to use subagents, but the presence of SKILL.md does not mean three independent agents ran. A marketplace is a catalog of plugins, not an audit or safety certification. Check the actual package contents before installation.

Public Claude Code SEO skills at a glance

These are three open-source starting points, not interchangeable audits. The component counts and scope below come from each project's README, checked 19 October 2026. We did not install or test the other two projects.

Choose by the job, then inspect the actual files and permissions before installing anything. A listed agent, script or data connection is a reason to review more of the package, not evidence that it ran in your project.

Before you install: inspect the public skill

For a concrete example, we inspected aevans-eng/seo-skill, a small public skill for static portfolio sites. Its MIT license permits reuse under its terms. GitHub's public commit metadata showed one release commit dated 14 March 2026 when checked on 19 October. The repository says it has one instruction file, no agent package, API calls or external dependency.

We tested this one because its single instruction file has no dependencies and can be audited in full before a first trial. That choice says nothing about how the larger packages perform.

Its SKILL.md asks Claude to inspect HTML for titles, descriptions, canonicals, social metadata, structured data, robots rules, sitemaps, images and links, then report a pass or fail table. This is a useful scope for a small static site. It is not a measured keyword-research service, a broad crawl of the live web, or the viral three-agent harness mentioned in some short videos. Those are different claims about different setups.

One reason to read the file first: it gives rigid length and markup suggestions. Google says search snippets are mostly drawn from page content, and there is no fixed meta-description length that Google always displays. A checklist may flag a description by character count; that is a prompt to read the description, not proof of a Google error. Google's structured-data guide likewise makes eligibility and content accuracy more important than adding markup indiscriminately.

How to install and try one safely

Anthropic documents project skills under .claude/skills/<name>/SKILL.md and personal skills under the user's Claude skills directory. For a one-page trial, prefer the project location in a disposable sample repository so you can see every file the skill touches. The public example's README provides its own setup directions, but inspect the file on GitHub first. Do not run a stranger's installation script merely because a reel says it is free.

1. Open the repository's README, license, instruction file and recent commits. Check whether it also contains hooks, scripts, agents or requests for credentials. Record the exact revision you reviewed.
2. Copy only the approved skill file to a fresh sample project's skill folder. Keep a Git checkpoint and do not provide production credentials or a client repository for a first run.
3. Ask Claude to audit one specified HTML page and produce findings without editing. Keep the page path and exact request in your notes.
4. For each finding, inspect the actual HTML and open the corresponding official guidance. A missing canonical, for example, is a different problem from a canonical that points to the wrong URL.
5. If you accept a change, request one change at a time, inspect the diff and run the project's own checks. Record the before page and the after page.

Anthropic's permissions guide describes how to limit tool use. A SKILL.md file is still an instruction source with access to whatever tools the Claude session is allowed to use. Its claims about being “safe,” “read-only” or “production-ready” are not permission to trust it with tokens or a live site.

What happened in our one-page test

On 19 October 2026, we placed the reviewed skill in an isolated sample project and asked Claude Code to audit only a 16-line HTML page. We allowed file reading, not edits, commands or web access. The page deliberately omitted a meta description, canonical link, social metadata and image alt text. It linked to a classes.html page and a workshop.jpg image that did not exist in the sample project.

Claude identified the missing description, canonical, social tags and image alt text, citing the relevant lines. It treated the skill's preferred title length as a recommendation rather than a search-engine requirement. It also recognized that the skill's suggested Person schema was a poor default for a woodworking workshop. This is the useful part of the test: a skill can organize a file-level review, while the operator still decides which recommendations fit the page.

The broken link and missing image file were not confirmed. Our prompt limited Claude to index.html, so it reported that it could not check whether those targets existed. The test therefore passed the metadata and restraint checks but did not prove link validation. Claude made no edits, and the HTML file's SHA-256 hash was unchanged before and after. We did not test a live site, rankings, indexing or AI citations, and the result should not be read as evidence for any of them.

What can the skill check, and what stays human?

For a static page, the reviewed skill can direct Claude to look for missing metadata, an inaccessible image description, a missing sitemap reference or inconsistent links. Those are inspectable file-level findings. The skill cannot tell from local HTML alone whether Google indexed the URL, which searches bring qualified visitors, or whether an AI answer later cites it. Those require separate evidence.

Treat robots rules as crawler instructions, not a way to force indexing. Treat sitemaps as discovery hints, not ranking guarantees. Structured data should describe visible content accurately. Check a proposed change against the page the visitor actually receives, including its mobile render, rather than accepting a plausible edit to a source file that the site does not publish.

The common failure is not that Claude cannot spot a missing tag. It is that a generic checklist may mistake a deliberate site choice for a defect, apply the same template to every page, or change a field without understanding the page's intent. If the skill proposes rewrites, ask what user question each rewrite answers and what evidence supports the new claim. Reject invented numbers, unsupported FAQs and schema for content that is not on the page.

What the social clips leave out

Recent short videos frame SEO skills as “harnesses” and suggest that adding several subagents produces a complete audit or an overnight optimization run. Anthropic's documentation confirms that skills, plugins and subagents are distinct building blocks. It does not establish that a particular public repo runs three agents or that a report is accurate. The inspected one-file repo makes a narrower promise: a static-site checklist.

An overnight run is a scheduling and permission decision, not a special property of an SEO skill. Before allowing unattended edits, define the repository, allowed tools, maximum spend, review point and rollback path. If an external tool supplies search-volume, rankings or Search Console data, name that tool and confirm access. A skill without those inputs should not present estimated figures as measured demand.

Measure whether the page earns AI citations

After the page is published, choose the exact question it answers and save a dated AI answer with its cited URLs. Repeat that question after the change. Xtrusio checks a published piece against the question it targets and counts it as verified only when a later answer cites that exact URL, as described in our AI visibility guide. A brand mention or a citation to a different page is a different observation. This measurement does not prove the skill caused the citation, but it shows whether the article achieved the outcome the team wanted to inspect.

For the example static page, record the initial HTML and the skill's proposed edits, publish only accepted changes, then compare the same question and exact citation later. Keep the date and engine with each answer. If no citation appears, check whether the page actually answers the question, is accessible, and contributes something the current cited pages do not. Do not add markup merely to chase a citation count.

Sources and scope

Checked 19 October 2026: Anthropic skills, plugins, marketplaces, permissions; the READMEs for aevans-eng/seo-skill, AgriciDaniel/claude-seo and factive1/claude-code-seo-skill; Google Search Central on snippets, structured data, robots rules and sitemaps.