Claude SEO skills: what they are, plus 5 SKILL.md files to copy

1. What Claude skills are, and where they run.

A skill is a folder with one required file, SKILL.md. The file opens with a short header between two lines of three dashes, holding a name and a description. Everything under the header is plain instructions, written the way you would brief a contractor. Under the Agent Skills specification the name runs to 64 characters of lowercase letters, numbers and hyphens and must match the folder name, and the description runs to 1,024 characters and should say both what the skill does and when to use it. A skill can also carry supporting files, such as a reference list or a script, that Claude opens only when the instructions point to them. Skills are one of the three setups in our Claude for SEO guide. This page goes deeper on that one setup and publishes the files.

The design that makes skills cheap to keep around is progressive disclosure. Per the Claude Code documentation, only each skill's name and description sit in Claude's context on every turn, and the body loads when the skill is used. You can keep a dozen SEO skills installed and pay for the one a task needs. The specification recommends keeping SKILL.md under 500 lines and moving long reference material into separate files, which is also the habit that keeps a skill readable when it needs fixing six months later. The same discipline (short files, clear sections, answer first) is what an llms.txt file asks of a website.

Where skills run:

• Claude Code. A personal skill lives at ~/.claude/skills/<name>/SKILL.md and works in every project on your machine. A project skill lives at .claude/skills/<name>/SKILL.md inside the repository, so committing it gives the whole team the same skill. Claude uses it when a request matches the description, or you run it directly as /name, and anything typed after the command reaches the skill as $ARGUMENTS.
• The Claude apps. In claude.ai and the desktop app you zip the skill folder and upload it under Customize, then Skills. Claude's Help Center says skills need code execution switched on, and custom skills you upload stay private until you share them.
• Plugins. A Claude Code plugin bundles several skills, plus optional subagents, hooks and MCP servers, into one unit you install from a marketplace. That is what most listings sold as a Claude SEO plugin are. A skill inside a plugin runs as /plugin-name:skill-name, and the Claude Code docs say plainly that whatever a plugin runs, it runs with your user privileges.
• Other agent tools. The format is published as an open standard, Agent Skills, and other agent tools have adopted it, so a well-written SKILL.md is not locked to one product.

Skills are not connectors. A skill teaches Claude a process. An MCP server hands it tools and data, such as your Search Console queries. The two combine well: a skill that says “pull the last 28 days of queries, then cluster them” is only as good as the connector behind it. The plumbing is covered in our guide to how an SEO MCP server works, and the roundup of the best MCP servers for SEO covers which ones to connect first.

2. How to write an SEO skill that fires at the right time.

Most skills that disappoint fail at the header, not the instructions. The description is what Claude reads when it decides whether a skill applies, so write it as the situation rather than the feature: what the skill does, then “Use when” and the words people actually type, such as audit, indexing, outline or schema. “Helps with SEO” matches everything and therefore nothing. The same rule holds for the instructions underneath: they are a brief, and a vague brief gets the same vague output from Claude that it gets from a person, which is the pattern our ChatGPT for SEO guide plans around for the other assistant.

Five rules separate the skills worth keeping from the ones that get deleted:

• One job per skill. A single skill called seo that audits, clusters and writes fires on every request and does each job badly. Five small skills each fire on their own trigger.
• Put the data rule in writing. A skill carries process, not data. There is no keyword database inside it, and Claude fills a missing figure with a guess that reads like a measurement. Every skill below forbids estimates in its rules section.
• Fix the output format. A table with the same columns on every run means this month's audit compares cleanly with last month's, and the output pastes straight into a ticket.
• Gate the side effects. In Claude Code, set disable-model-invocation: true on any skill that edits files or publishes, so it runs only when you type the command, and use allowed-tools to pre-approve the one command a read-only skill needs (the audit below pre-approves curl). Those fields are Claude Code features. The name and description are the part every surface reads.
• Read before you install. Claude's Help Center tells you to install skills only from trusted sources and to review the files first, because a skill can instruct Claude to run code or contact outside services.

To install the five skills below in Claude Code, save each one as SKILL.md inside a folder named exactly like its name field, for example .claude/skills/seo-site-audit/SKILL.md, then run it as a slash command followed by its arguments. For the Claude apps, zip each folder and upload it. Each skill stands alone, and the checks inside them come from the same lists we hold our own pages to, starting with the SEO checklist we publish.

3. Skill 1: a technical site audit.

The audit skill works from what the site actually serves: robots.txt, the XML sitemap and a sample of pages fetched with curl. That is why it is built for Claude Code, where Claude can run commands. It checks the things that stop a page being indexed before anything cosmetic, and every line traces back to our technical SEO audit checklist, trimmed to what a single session can fetch.

---
name: seo-site-audit
description: Technical SEO audit of a live website from its robots.txt, XML sitemap and a sample of pages. Use when asked to audit a site, check indexability, or find out why pages are not indexed or not showing in search.
argument-hint: [domain]
allowed-tools: Bash(curl *)
---

# SEO site audit

Audit the site at $ARGUMENTS. Work only from what the site serves.

## Steps

1. Fetch https://$ARGUMENTS/robots.txt. Record every Disallow rule,
   every Sitemap line, and any rule that blocks Googlebot or the AI
   crawlers GPTBot, OAI-SearchBot, ClaudeBot, Claude-SearchBot,
   PerplexityBot or Google-Extended.
2. Fetch each sitemap listed there (fall back to /sitemap.xml). Follow
   sitemap index files one level deep. Count the URLs.
3. Pick up to 25 URLs: the homepage, the 10 most recent by lastmod,
   and the rest spread across sections of the site.
4. Fetch each URL with curl -sL and record:
   - final status code and any redirect chain
   - title tag and its length in characters
   - meta description and its length
   - meta robots and X-Robots-Tag values
   - canonical URL, and whether it points to the page itself
   - number of H1 tags
   - JSON-LD types present
5. Flag every sitemap URL that returns a non-200 status, redirects,
   is noindex, or canonicalises to another URL.

## Output

A table, one row per URL:
URL | status | title length | meta length | canonical | H1 count | robots | schema

Then findings sorted by severity: blocks indexing, wastes crawl,
cosmetic. Each finding names the URLs affected and the exact fix.

## Rules

- Report only what you fetched. If a fetch fails, say so and move on.
- Never estimate traffic, rankings or search volume. You have no data
  for them.
- Do not edit any file. This skill reports. Fixes are a separate step.

Run it as /seo-site-audit yoursite.com. A 25-page sample keeps one run inside a normal session. Raise the number in step 3 for a bigger site, or run it once per section of the sitemap. The robots.txt step names GPTBot, ClaudeBot and the other AI crawlers because a rule written for one bot often blocks another, and blocking them is a business decision rather than a default (the trade-off for each Claude bot is in what ClaudeBot does and how to control it). Source code and fetched HTML can still differ from what a browser renders, so compare the result with our free SEO audit of the live site.

4. Skill 2: keyword clustering.

Clustering is where Claude is strongest and where it is most tempting to let it make things up. Given a list, it groups by intent rather than shared words, which a spreadsheet filter cannot do. Asked for the list's numbers, it invents them. This skill takes a CSV (a Search Console query export, or an export from whichever keyword tool you pay for) and uses only the numbers in the file. If you have no list yet, our free blog post ideas generator gives you seed topics to expand.

---
name: keyword-clustering
description: Groups a keyword list into page-level clusters by search intent and maps each cluster to an existing or new URL. Use when given a keyword export, a Search Console query list or a CSV of terms to cluster, dedupe or map to pages.
argument-hint: [path to keyword CSV]
---

# Keyword clustering

Cluster the keywords in $ARGUMENTS (a CSV file or a pasted list).

## Input

Expected columns: keyword, plus any of volume, difficulty, clicks,
impressions, position. Use only the numbers in the file. If a column
is missing, leave it blank. Never fill it in.

## Steps

1. Normalise: lowercase, trim, and merge exact duplicates and plural
   or word-order variants ("seo audit tool", "seo audit tools").
2. Label each keyword's intent: informational, commercial,
   transactional or navigational.
3. Group keywords that one page could rank for: same intent and same
   core topic. If two keywords need different page formats (a list
   vs a how-to), they go in different clusters even if they share words.
4. Name each cluster by its head term: the keyword with the highest
   volume in the file, or the clearest short phrasing if there is no
   volume column.
5. If a sitemap or URL list is provided, map each cluster to the URL
   that already targets it, or mark it NEW.
6. Flag cannibalisation: two existing URLs mapped to one cluster.

## Output

A table:
cluster | head term | intent | keyword count | total volume (from file) | URL or NEW

Then the 5 NEW clusters to build first, ranked by total volume from
the file, with one line each on why.

## Rules

- Never estimate search volume, difficulty or CPC. With no numbers in
  the file, say the ranking is by judgment only.
- Every keyword appears in exactly one cluster.
- Keep each cluster small enough for one page. Split when in doubt.

Hand it a sitemap or URL list along with the CSV and step 5 turns the clusters into a page map: which existing URL owns each cluster, which clusters need a new page, and where two of your pages already compete for the same one. That last flag is the cheapest win in the output, and fixing it is the same job our blog SEO guide starts with. When the file has no volume column, the skill says the ranking is judgment only. That is the honest answer, and it is the gap a skill cannot close on its own.

5. Skill 3: a content brief.

A brief is the step that decides whether a draft is worth editing. This skill reads the pages that rank, pulls the questions searchers ask next, and returns an answer-first outline with every fact tied to the page it came from. It needs web search, so run it where search is available and open the sources it cites. If you draft from the brief in Claude, the largest model is the right call for this job, and our guide to the best Claude model for writing compares the current lineup.

---
name: seo-content-brief
description: Writes an SEO content brief for one target keyword from the pages that currently rank for it. Use when asked for a content brief, an article outline, or a plan for a new page that targets a keyword.
argument-hint: [target keyword]
---

# SEO content brief

Write a brief for the keyword: $ARGUMENTS

## Research

1. Search the keyword. Open the top 5 organic results that are
   articles or guides. Skip forums, videos and product pages unless
   that is all that ranks, and say so if it is.
2. For each page, note its H2s, its format (how-to, list, comparison,
   definition), its rough length, and what it answers in its first
   100 words.
3. Note the follow-up questions searchers ask (People Also Ask,
   related searches) where the results show them.

## The brief

- Target keyword, plus 3 to 6 secondary phrasings taken from the
  pages and questions you read.
- Search intent in one sentence, and the format the results reward.
- The direct answer: 2 to 4 sentences that answer the query, to open
  the article.
- H2 outline: 6 to 9 headings in order, each with one line on what it
  must cover.
- Gaps: what every ranking page misses that this article will cover.
- Facts to source: each statistic or claim the article needs, with the
  page you found it on. Anything without a source is marked
  SOURCE NEEDED.
- Internal links: 3 to 5 pages from the site's URL list (if provided),
  with the anchor text for each.
- Length: a range based on the ranking pages, not a fixed number.

## Rules

- Never invent statistics, studies or quotes. A claim with no source
  stays marked SOURCE NEEDED.
- Do not state search volume or difficulty unless it was given to you.
- Do not write the article. The brief stops at the outline.

Two rules do most of the work. SOURCE NEEDED keeps an unsourced number visible instead of letting it pass as fact, and “do not write the article” keeps the brief short enough to review in two minutes. Every SOURCE NEEDED marker is still a claim somebody has to check by hand before publishing, which is the job an automated fact-checking pass exists to take off your desk.

6. Skill 4: on-page checks.

The on-page skill checks one page against ten items and returns paste-ready rewrites. It takes two arguments, the page and the target keyword, which Claude Code passes in as $0 and $1. Arguments split on spaces, so wrap a multi-word keyword in quotes: /on-page-seo-check yoursite.com/blog/my-post "claude seo skills". Use it on a draft before it publishes, and on a live page sitting on page two, where a sharper title is often the cheapest fix available. Our free SEO title generator does the title rewrite alone when that is all a page needs.

---
name: on-page-seo-check
description: Checks one page against an on-page SEO list (title, meta description, headings, keyword placement, links, images, canonical, schema) and returns paste-ready fixes. Use when asked to check, review or optimise a single page, post or draft for SEO.
argument-hint: [URL or file path] [target keyword]
---

# On-page SEO check

Check the page at $0 for the target keyword $1.

If $0 is a URL, fetch it. If it is a file, read it. If no target
keyword was given, ask for one before starting.

## Checks

1. Title tag: present, unique, keyword or a close variant near the
   start, roughly 50 to 60 characters.
2. Meta description: present, matches the page, written for the
   click, roughly 140 to 160 characters.
3. Exactly one H1 that states the topic. H2s and H3s in order, with
   no skipped levels.
4. The first 100 words answer the query directly.
5. Keyword in the H1, the first paragraph and at least one H2, read
   naturally. Flag stuffing.
6. At least 3 internal links to relevant pages, descriptive anchors,
   no "click here".
7. Every content image has alt text that describes it.
8. URL is short, lowercase, hyphenated and names the topic.
9. Canonical present and self-referencing. Page is not noindex.
10. Structured data present for the page type and matching the
    visible content.

## Output

PASS, FIX or FAIL for each check, with the evidence (the actual title,
its counted length, the actual H1). Then the rewrites, ready to paste:
a new title, a new meta description and any heading changes.

## Rules

- Quote what is on the page. Do not assume.
- Count characters. Never estimate a length.
- Never claim a change will lift rankings by a specific amount.

Counting is the part to trust least in any language model, so the skill tells Claude to count rather than estimate, and you should still spot-check the numbers. For single fields the free tools are faster: the meta description generator builds the length limit in, the keyword density checker settles a stuffing argument in seconds, and the internal linking tool maps the links that check 6 asks for.

7. Skill 5: schema markup.

Schema is quick for Claude to write and easy to get wrong in one specific way: marking up things the page does not show. This skill takes every value from the visible page, writes one JSON-LD block per type, and lists the fields it left out and why. Structured data helps search engines and answer engines read a page, and the answer engine optimization guide covers how those engines pick what to quote.

---
name: schema-markup
description: Writes JSON-LD structured data for a page from its visible content (Article, BlogPosting, Product, Organization, LocalBusiness, BreadcrumbList) and checks existing markup against the page. Use when asked to add, fix or validate schema markup or structured data.
argument-hint: [URL or file path]
---

# Schema markup

Write or fix the JSON-LD for $ARGUMENTS.

## Steps

1. Read the page and pick the main type: Article or BlogPosting for a
   post, Product for a product page, Organization or LocalBusiness
   for a homepage or about page. Add BreadcrumbList if the page shows
   breadcrumbs.
2. Take every value from the visible page only: headline, author,
   dates, image, price, availability, address, opening hours.
3. Write one JSON-LD block per type, inside a script tag with
   type="application/ld+json".
4. If the page already has markup, compare it with the visible
   content and list every field that is missing, wrong, or not shown
   on the page.

## Output

The JSON-LD, ready to paste. Then a list of the fields you left out
because the page does not show them.

## Rules

- Mark up only what a visitor can see. Never add ratings, reviews,
  prices or FAQs that are not on the page.
- Use absolute URLs.
- Do not promise rich results. Google decides eligibility, and it has
  restricted or retired several rich result types.
- Tell the user to test the output in Google's Rich Results Test and
  the Schema Markup Validator before it ships.

Markup does not make a thin page rank. Pair it with the answer-first structure that LLM SEO depends on, test every block in Google's Rich Results Test and the Schema Markup Validator before it ships, and treat any rich result as a bonus that Google may or may not grant.

8. Where Claude SEO skills stop.

Five skills cover a real share of SEO work, and for one site at a time they are the right amount of tooling. Their limits come from what they are. A skill is instructions, so it knows no search volume and sees no rankings unless a connector supplies them (the MCP servers that bring SEO data into Claude fill part of that gap). It runs when you ask, so nothing happens on the days you do not. And its output lands in a chat or a file, so the fixes, the drafts and the publishing are still yours to do.

Stack the five into a publishing program and the gaps show. Every article needs a keyword picked from measured data, a check against the pages you already have, a brief, a draft, a fact-check, a fix pass, internal links, schema and a publish to your CMS. That loop is what people mean when they search for a Claude SEO agent. You can build one from skills, MCP connectors and a scheduler, but you own the glue, the credentials and the morning a step fails without telling anyone.

That loop is what The SEO Agent runs. It picks keywords from real keyword data (live monthly volume and difficulty, never an estimate), checks your existing pages so it never duplicates one, runs a fact-check pass with a cited source for each claim, refuses drafts that fail its quality gate, links only to URLs that exist, and publishes natively to WordPress, Webflow, Shopify, Wix, Ghost or Framer every day. The full SEO automation pipeline shows each stage, it costs $99 a month after a free trial, and if you live in Claude Code you can call the same pipeline from Claude over MCP instead of wiring each skill yourself.

A worked example: installing the audit skill and reading its first run.

Here is the whole path for Skill 1 in Claude Code, from an empty folder to a findings list. The site is a placeholder, example.com, and the output below is illustrative: it shows the shape a run returns, not a real audit of a real site. From the root of your repository:

mkdir -p .claude/skills/seo-site-audit
# save Skill 1 as .claude/skills/seo-site-audit/SKILL.md
claude
/seo-site-audit example.com

Claude reads robots.txt, walks the sitemap, samples 25 URLs and fetches each one. The curl calls run without a prompt because the allowed-tools line pre-approved them, and nothing else does. The report comes back in the format the skill fixed:

URL               status  title  meta  canonical  H1  robots   schema
/                 200     54     151   self       1   index    Organization
/pricing          200     71     0     self       1   index    none
/blog/old-guide   301     n/a    n/a   n/a        n/a n/a      n/a
/blog/new-guide   200     48     138   /blog/     2   index    Article
/tag/updates      200     22     0     self       0   noindex  none

Blocks indexing
1. /blog/new-guide canonicalises to /blog/, so search engines are told
   to index the blog home instead. Fix: point the canonical at the
   page's own URL.
2. /blog/old-guide is in the sitemap but redirects (301). Fix: list
   the destination URL in the sitemap instead.

Wastes crawl
3. /tag/updates is noindex but listed in the sitemap. Fix: remove it
   from the sitemap.

Cosmetic
4. /pricing: title is 71 characters, and there is no meta description.
5. /blog/new-guide has two H1 tags.

robots.txt: GPTBot is disallowed site-wide, ClaudeBot is allowed.
Confirm this is intended.

What makes this useful is the ordering and the evidence. The two findings that stop pages being indexed come first, each with the URL and the fix, and nothing in the report is a guess: lengths are counted, statuses are fetched, and the skill says nothing about traffic because it has none. The canonical fix is a ten-minute job. Knowing which pages matter most needs Search Console data, which is where a connector comes in.

What the run cannot tell you is the next step. It found five problems on 25 pages. It did not fix them, did not check the pages it skipped, and will not run again next week unless you run it. Make the fixes on a branch, check the rendered pages rather than the templates, and then run Skill 4 on the pages that earn the most clicks.

Common mistakes with Claude SEO skills.

1. One skill called seo that does everything. It feels efficient to write one file. In practice its description matches every request, it loads a wall of instructions for a one-line question, and each job inside it gets a fraction of the attention. Split it into one skill per job.
2. A description that names the feature, not the situation. “SEO audit skill” tells Claude what the file is, not when to use it. Write what it does, then “Use when” and the words a request actually contains.
3. Letting a skill fill in numbers. Without a rule, Claude will estimate search volume, difficulty or traffic, and the estimate reads exactly like a measurement. Put the ban in every skill and bring numbers in from an export or a connector.
4. Installing skill packs and plugins you have not read. A plugin runs with your user privileges, and a skill can tell Claude to run code or contact outside services. Read every SKILL.md and script first, and prefer short skills you can audit in a minute.
5. Auto-running skills that change things. A skill that edits templates or publishes posts should never fire because a sentence happened to match its description. Set disable-model-invocation on it, work on a branch, and read the diff before you merge.
6. Treating one run as monitoring. An audit is a snapshot of 25 pages on one afternoon. Skills do not schedule themselves, so a broken canonical shipped next Tuesday stays broken until somebody runs the skill again.