claude-seo (Universal SEO skill for Claude Code)

A Claude Code plugin packing 25 sub-skills, 18 specialist sub-agents and 32 commands for SEO auditing, by agrici-daniel (MIT, 14.8k★ / 2.2k forks, 203 commits, v2.2.4 July 2026, single maintainer). Shared via Telegram, ingested 2026-08-22.

It is the second instance of the SEO-skill-pack form in this spoke and by far the larger: agentic-seo-skill (2026-05-29) shipped 16 sub-skills, 10 agents and 89 scripts. The same placement rule applies to both — the SEO domain is incidental, the skill-pack form factor is why it is here. The domain substance is cross-spoke context for ../search-marketing-wiki, recorded below rather than split out.

The shape

Three layers, with the orchestration layer doing the work:

Directive — skill metadata and entry points (skills/seo/SKILL.md is the orchestrator).
Orchestration — auto-discovery from skills/seo-*/ and agents/seo-*.md, industry detection (SaaS / local / ecommerce / publisher / agency), then parallel dispatch of up to 15 sub-agents.
Execution — specialists return findings, the orchestrator synthesizes.

Auto-discovery is what separates this from a folder of prompts: adding skills/seo-newthing/SKILL.md registers a capability without editing an orchestrator. That is agent-orchestration's fan-out with a convention-based registry underneath, and it is the mechanism a 25-skill pack needs to stay editable by one person.

Install is through the plugin marketplace (/plugin marketplace add AgriciDaniel/claude-seo), with shell and PowerShell fallbacks. Outputs are Markdown, JSON, and WeasyPrint A4 PDFs (~32 pages for a full audit, matplotlib charts at 200 DPI). State lives in ~/.config/claude-seo/, credentials at 0o600.

Two things here the corpus has not seen in a skill pack

It ships a real test suite, and most of it is security. 410 tests, up from 39 at v1.9.9, with 91 cases covering URL-safety SSRF and DNS-rebinding bypasses — obfuscated IPv4, FQDN trailing-dot, redirect-rebinding. A CI workflow runs them. Every other skill pack this spoke holds is markdown with no executable contract; this one treats its fetcher as attack surface, which it is: the tool takes a URL from a user and fetches it from the machine running the agent. Two releases (v2.2.0) are credential- injection and SSRF-bypass fixes, so the risk is not hypothetical to the author either.

Each recommendation carries its own disconfirmation. The stated framework is 10 principles across four phases — PERCEIVE (observe, listen) → ANALYZE (think, connect) → VALIDATE (feel, accept) → ACT (create, grow) — and the documented output contract is that every recommendation ships "the first-principle observation it rests on, its dependency relationship to other recommendations, a 'how would we know this failed?' check, and a leading indicator."

That framework is a cognitive skill in reasoning-scaffolds' sense: a named sequence prescribing how to think, adding no capability the model lacked. The spoke's standing finding is that cognitive scaffolds are the kind with measured and negative evidence (skill-induced-failure, reasoning-scaffolds). What is new is the falsifiability clause. A scaffold that requires each output to state how it could be wrong is the first one here whose own product is checkable — it does not show the framework works, but it makes the recommendations testable, which is a different and rarer move than asserting rigour.

Opinionated positions, and one lands on another spoke

The pack takes named stances rather than staying neutral:

Rejects three AI-search myths — llms.txt as a citation lever, a content-chunking requirement, and AI-specific keyword rewriting — aligning to Google's May 2026 AI Optimization Guide, with the reasoning kept in skills/seo-geo/references/llmstxt-evidence.md.
E-E-A-T with Trust weighted highest, per the September 2025 Quality Rater Guidelines, YMYL widened to political and social topics.
JSON-LD preferred over Microdata/RDFa; FAQ rich results treated as retired (2026-05-07) while QAPage stays valid for genuine Q&A; a deprecated-types-2024-2026.md reference tracks the churn.
Headless rendering by default (Playwright + trafilatura), auto-detecting Next.js, React, Vue, Nuxt and Astro.

Cross-spoke: ../search-marketing-wiki already holds the stronger version of the llms.txt finding — llms-txt records Google's own statement that "Google Search itself doesn't use them". So this source adds no new evidence there. What it does add is propagation: a 14.8k-star practitioner tool has encoded that finding as a default and written the reasoning into a reference file its agents read. A corpus finding reaching shipped tooling is the signal here, not the claim itself. Recorded as a pointer, not as a second page.

Reach, and what it costs

Eight optional MCP extensions (DataForSEO, Firecrawl, Ahrefs' official server, SE Ranking, Profound, Bing Webmaster + IndexNow, Unlighthouse, Banana) and a four-tier Google credential ladder — Tier 0 API key for PageSpeed/CrUX, Tier 1 OAuth for Search Console, Tier 2 GA4, Tier 3 Ads for Keyword Planner. Its own framing of the competition is a boundary rather than a claim of superiority: Screaming Frog is a crawler, Ahrefs is a backlink index, and claude-seo wraps Ahrefs over MCP rather than replacing it.

What is not evidence here, stated plainly

The "10–15 minutes versus 4–8 hours manual" figure is the repo's own, with no method behind it.
The case study is one screenshot. A single site launched 2026-03-23 showing rising Search Console impressions through 2026-06-12. Three months, one site, no control, no attribution — that is a testimonial, not a benchmark. The 410 tests measure the code; nothing here measures whether the audits improve rankings.
Free tier is local-only. No third-party calls by default, so Core Web Vitals are lab estimates until Google credentials are supplied, and backlink and competitive data need paid MCP extensions.
Author-stated limitation: pages with scroll-bound hydration or post-interaction fetches produce noisy findings, with a manual seo-visual cross-check as the workaround.
Single maintainer at 14.8k stars — the bus factor this spoke keeps noticing on popular skill packs.

The distribution pattern, recorded

The MIT repo has a private mirror (AI-Marketing-Hub/claude-seo) giving early access behind a Skool membership, alongside a free and a paid community, a YouTube channel and a blog. So the open-source pack is also the top of a paid-community funnel. That is neither hidden nor unusual, and it is worth naming because the spoke tracks how skill packs are distributed: this is the first here with a two-tier open/paid release cadence.

Name-collision caution: the README lists an interoperating skill called "AI Marketing Claude". This spoke already holds ai-marketing-claude by a different author, zubair-trabzada. They may or may not be the same project; nothing read here settles it, so the two are kept separate and this is flagged rather than merged.

No longer the only funnel (2026-08-31). geo-seo-claude (zubair-trabzada, MIT, 9,877★) arrived with the same shape: repository homepage pointing at a Skool community, and a README section naming the business the free pack feeds — "GEO agencies charge $2K–$12K/month. This tool does the audit. The community teaches you how to sell it." Different structure, though: no private mirror and no early-access tier, so the code is not two-tier at all. Two of the three largest SEO packs here are now the top of a paid Skool funnel, by two unconnected authors, which makes this a pattern in how marketing skill packs are distributed rather than one maintainer's arrangement. What stays specific to this pack is the open/paid release cadence — gated access to the same code — which the other does not do.

agentic-seo-skill · ai-marketing-claude · claude-skills-ppc · agent-skills · agentskills-spec · agent-orchestration · reasoning-scaffolds · skill-induced-failure · model-context-protocol · claude-code · agrici-daniel · zubair-trabzada · geo-seo-claude · llms-txt · e-e-a-t