Claude SEO skills: catalog, use, and build

Claude SEO skills are reusable job packages - a SKILL.md plus optional refs/scripts - not a prompt dump and not a marketplace shopping list. Match skill types to weekly SEO jobs (briefs, audit triage, on-page QA, schema, internal links, GEO-light), load SERP/crawl/brand inputs in a Project first, then install or author when the loop is stable. Install order: What Claude SEO skills to install first. GitHub find/install: How do I find and install a Claude SEO skill from GitHub?. Jobs playbook: Claude for SEO. Twin: Claude marketing skills.

When to use Claude SEO skills / when not

Use skills when you repeat an SEO job weekly with stable inputs: SERP notes, crawl or GSC extracts, brand voice, URL list, and a clear QA bar. Same brief or audit shape every run → skill territory.

Don't use skills yet on day zero with empty data, one-off spikes, strategy/taste calls, or a "40-skill agency pack." Skills don't invent rankings or Labs volumes. Overlapping descriptions (audit vs page vs technical) compete for trigger space - more skills can make the right one less likely to fire.

Install order only → install-first. GitHub discovery/install → find-install. How to run audits/briefs/ops → Claude for SEO.

Catalog map - skill types by SEO job

Catalog by job, never by unverified marketplace SKU names. Third-party SEO skill repos exist - verify name, description, and license same-day; install path on the find-install guide. Rows are types, not product titles.

Do not invent live marketplace titles. Twin: Claude marketing skills.

Workflow - use or build in 8 steps

1. Inventory weekly SEO jobs - What you actually run (≥1×/week). Align to Claude for SEO; don't rebuild the playbook.

2. Load a Project pack - Brand voice, ICP, proof, SERP/crawl/GSC notes, URL list. Skills without context still produce generic filler.

3. Run as a prompt until the loop is boring - Role · inputs · steps · rules · output. Structure parallel: Claude prompts for marketing.

4. Choose the surface - Chat for one-offs; Projects for brand + SERP + repeat briefs; Claude Code for repo/scripts/agents; marketplace only after same-day name/description/license check.

5. Install or author - Matching skill or thin custom SKILL.md when the prompt barely changes. Order → install-first; GitHub → find-install; marketplace pattern → marketplace add.

6. Test the trigger - Teammate phrases the job three ways ("audit this URL" vs "technical SEO check"); description must match real speech.

7. QA the output - No invented rankings, Labs volumes, or competitor facts; evidence fields required; human gate for customer-facing pages.

8. Prune - Remove unused skills. Triggers fail or context bloats → skill-doctor.

Prompts (≥3 copy-paste)

Three specs for skill decisions - not the SEO jobs prompt pack. Mirror: prompts.md. Job prompts live on Claude for SEO.

1. SEO skill-need diagnostic

Role: SEO ops lead. Inputs: job list · frequency · data available · prompt stability. Output: skill / prompt / defer table. When not: day-zero with zero SERP/crawl/brand inputs.

You are an SEO ops lead. Decide which jobs deserve a Claude skill vs stay as prompts.

REQUIRED INPUTS: Weekly SEO job list (name + frequency); Data available (SERP notes / crawl-GSC / brand voice / URL map or "none"); For each job: prompt stable? (yes/no/unknown); Surface preference (chat|Projects|Claude Code|marketplace|unknown).

STEPS: (1) Score each job: repeat weekly? has inputs? stable loop? clear QA bar? (2) Table: Job | Verdict (skill|prompt|defer) | Why | Next action. (3) Cap "skill" recommendations at 3 unless every row clearly passes. (4) Never invent marketplace skill product names - use job-type labels only (brief, audit, QA, schema, linking, GEO-light, outline, research). (5) Route "what to install first" to the install-order companion; route GitHub install how-to to the find-install companion; route audit/brief playbooks to Claude for SEO.

RULES: Fail any score row → prompt or defer. No "install everything SEO" advice. Mark [GAP] where inputs are missing. No invented Labs volumes or rankings.

OUTPUT: decision table + 3 next actions + QA self-check (no invented skill names, cap respected, gaps marked).

2. SKILL.md draft from a stable SEO prompt

Role: Skill author. Inputs: winning prompt · job name · trigger phrases · optional refs. Output: frontmatter + instructions body. When not: prompt still changing weekly.

You are a Claude Agent Skills author. Convert one stable SEO prompt into a thin SKILL.md.

REQUIRED INPUTS: Job name (kebab-case candidate, e.g. seo-page-brief); Winning prompt (full text); When to use (3-5 user phrasings); When not; Required inputs list (SERP/crawl/brand/URL); Optional reference filenames (or "none"); Surface (claude.ai|Claude Code|both).

STEPS: (1) Draft YAML frontmatter with name (lowercase kebab-case) and description that states WHAT + WHEN. (2) Body: Instructions (steps), Required inputs, Output schema (evidence fields), QA checklist, When not / stop rules. (3) Keep body lean; put long checklists in optional refs. (4) Flag any hard-coded algorithm or product "facts" that will expire - move those to Project files. (5) Do not add marketplace packaging claims or invent rich-result guarantees.

RULES: Description ≤1024 chars; name ≤64 chars, a-z/0-9/hyphens only. No invented rankings, volumes, or competitor weaknesses in examples. Label stub vs production if packaging is incomplete.

OUTPUT: complete SKILL.md in a fenced block + trigger-test checklist (3 alternate phrasings).

3. Trigger / description QA for SEO skills

Role: Skills QA. Inputs: draft description · sibling skill descriptions · sample requests. Output: pass/fail + rewrite. When not: no sibling list and no real teammate phrasing.

You are QA for Claude SEO skills. Stress-test a skill description for false triggers and misses.

REQUIRED INPUTS: Candidate skill name; Draft description; Sibling skill descriptions (or "none" - include audit vs brief vs schema if they exist); 5 real request phrasings from the team; Jobs this skill must NOT own.

STEPS: (1) For each phrasing: trigger? (yes/no/borderline) + why. (2) Flag overlap with siblings (especially audit vs page brief vs schema). (3) Rewrite description if misses or overlaps. (4) List phrases that should stay on prompts or route to Claude for SEO playbooks, not this skill. (5) Never invent competing marketplace skill titles.

RULES: Prefer precise WHEN language over buzzwords. One job per skill. Cap rewrite at 1024 characters.

OUTPUT: trigger table | overlap notes | revised description | ship/hold verdict.

Agent block - SEO skill librarian

Paste into Project instructions or a Claude Code custom agent. Job match, no invented marketplace names, routes install-order / GitHub / jobs questions to companions.

You are the SEO skill librarian for this brand.

RULES:
1. Map every request → SEO job type (brief, audit triage, on-page QA, schema, internal links/cannibalization, GEO-light, outline/ops, research, other).
2. Recommend skill OR prompt OR defer - never invent marketplace/plugin skill product names. Use job-type labels only.
3. If SERP/crawl/brand/URL inputs are missing, mark [GAP] and ask - do not invent fills, rankings, or Labs volumes.
4. Install-order / "what should I install first" → point to https://claude-for.com/guides/what-claude-seo-skills-to-install-first ; do not invent a day-one top-3 SKU list.
5. GitHub find/install how-to → https://claude-for.com/guides/how-do-i-find-install-claude-seo-skill-github ; do not paste a full install recipe here.
6. Audit / brief / content-ops playbooks → https://claude-for.com/guides/claude-for-seo .
7. Prefer Projects + prompts until the loop is stable; graduate to SKILL.md only when the prompt barely changes.
8. Customer-facing SEO output is always draft until a human QA gate.

OUTPUT FORMAT each turn:
- Job type
- Recommendation (skill type | prompt | defer)
- Next action (1 sentence)
- QA checklist (inputs present? one job? no invented metrics/rankings? human gate?)

Skills path / stub + SKILL.md anatomy

Honest stub: skill-stub/SKILL.md (seo-page-brief). Structure only - not a zip and not a marketplace listing.

Anatomy: YAML name + trigger-rich description (what and when); body = required inputs, steps, output schema with evidence fields, stop/QA rules; optional references/ for checklists. Official: Agent Skills · Claude Code: Extend Claude with skills · (Checked 2026-09-19).

Bridges: install-first · find-install · marketplace add · skill-doctor. No invented Anthropic "SEO skill" SKU.

Stack glance

Projects beat skills when context is the bottleneck; skills beat re-pasting when the procedure is stable. Map (1×): Claude for marketing.

Pitfalls + quality bar

• Skill bloat - overlapping audit / page / technical descriptions dilute triggers.
• Rules vs examples contradiction - QA skill bans invented metrics while examples invent CTR curves.
• Hard-coded expired "algorithm facts" - keep volatile claims in Project files, not the skill body.
• Listicles as product truth - third-party "25 SEO skills" repos are not Anthropic docs; verify same-day.
• Autopilot publish - skill output is draft until a human QA gate.
• Wrong spoke - catalog/use/build here; caps → install-first; GitHub commands → find-install; playbooks → Claude for SEO.

Quality bar: one job · sourced evidence · no invented Labs/rankings · on-brand · human QA for customer-facing pages.

FAQ

What are Claude SEO skills? Reusable SKILL.md packages (+ optional refs/scripts) for an SEO job - not a one-shot prompt, agent, or MCP. Metadata always loaded; body when triggered (Agent Skills overview, Checked 2026-09-19).

How are they different from "Claude for SEO" workflows? This page = catalog/use/build. Claude for SEO = audits, briefs, content ops.

Should I install SEO skills before writing prompts? Usually no - Projects + prompts until stable, then graduate. Order: install-first.

How do I find and install an SEO skill from GitHub? Marketplace/plugin when trusted, or copy a verified SKILL.md into your skills dir and smoke-test. Recipe: find-install.

How do I build an SEO SKILL.md? Frontmatter name + description (what and when), instructions, required inputs, output schema, optional refs. Test on a live URL/brief. Stub: skill-stub/SKILL.md. Prune: skill-doctor.

Key takeaways

• Catalog Claude SEO skills by job type, not marketplace fame.
• Load SERP/crawl/brand inputs in a Project; graduate stable prompts into thin SKILL.md files.
• Use chat · Projects · Claude Code · marketplace deliberately; verify plugin names same-day.
• Cap early installs; prune unused skills; never invent skill product titles.
• Install order → install-first; GitHub → find-install; jobs → Claude for SEO; twin → Claude marketing skills.

Sources (checked 2026-09-19)

• Agent Skills overview · Extend Claude with skills · Equipping agents for the real world with Agent Skills
• Live: Claude for SEO · install-first · find-install · Claude marketing skills · marketplace add · skill-doctor · Claude for marketing · prompts for marketing · marketing install-first
• Planned (not linked): claude-for-growth · #7 claude-code-for-seo · #8 claude-for-geo-aeo