The 480-repo marketing-skill cluster on GitHub, in plain terms

The agency-as-distribution genre we named in the ColdIQ audit ships ~12 GTM/sales repos. The marketing-side equivalent is the bigger half of the story by a wide margin.

A May 2026 sweep across sixty-plus targeted GitHub queries (filename:SKILL.md plus marketing terms, plus topic queries against claude-skills, agent-skills, cursor-rules, plus tree-walks of the major monorepos) surfaces 480-plus marketing-flavored skill repos. The top ten alone ship over five hundred individual SKILL.md files spanning copywriting, SEO, ads, social, email, CRO, analytics, brand, and ASO. The marketing-skill cluster is the largest, fastest-growing application of Anthropic’s SKILL.md spec since the format landed in late September 2025.

This piece walks the cluster. Four canon leaders. A bimodal quality distribution. Eight categories with real depth. Four gaps that haven’t been filled at the same intensity.

The four canon leaders

Star counts inside the marketing-skill cluster are inflated by viral cloning, but the leaders are real. Each ships meaningfully more than a single skill, has frontmatter discipline, and is being forked by other repos as a starting template.

The Corey Haines pack is the canonical reference. Its description fields are the cleanest trigger writing in the space, with phrases like “this copy is weak” and “make this more compelling” that map cleanly to what marketers actually say to themselves about a draft. Its six-skill funnel-stage CRO breakdown (page-cro, signup-flow-cro, onboarding-cro, paywall-upgrade-cro, popup-cro, form-cro) is the best-structured CRO taxonomy in the cluster.

The two AgriciDaniel repos sit underneath the Corey Haines pack as the depth references. SEO and paid ads are the categories where channel-specific knowledge compounds, and Daniel Agrici has built the canon for both. claude-seo ships twenty-four named sub-skills inside one monorepo, plus twelve specialist subagents, plus a Banana extension layer for LLM-assisted SERP work.

zubair-trabzada/geo-seo-claude at almost seven thousand stars is the AI-search-first SEO pack. It diverges from Agrici’s by focusing on citability inside ChatGPT, Perplexity, Claude, and Gemini answers (which is what GEO and AEO mean in practice) rather than Google blue-link rank. The fact that two distinct SEO packs have crossed five thousand stars in eight months is a signal about how much demand there is for production-grade marketing scaffolding.

thatrebeccarae/claude-marketing is the outlier worth naming. Twenty-six stars at the time of capture, against fifty-six versioned skills with tested: and tested_with: keys showing the author validated each skill against a specific Claude Code build. It’s the highest-meta-quality repo in the cluster and the most under-distributed. Operators evaluating skill packs for production use should weight commit hygiene over star count, and Rebecca Rae’s repo is the cleanest practical example of why.

The quality distribution is bimodal

The 480 number is a real count. The quality behind it isn’t evenly spread.

A hand audit of the top eighty repos by stars and recency lands on three buckets:

The long tail isn’t dishonest; it’s a natural artefact of an ecosystem where forking someone’s skill folder is the supported onboarding path. But star counts inside the cluster don’t separate the tail from the canon, and that’s the trap. coreyhaines31/marketingskills at 26,287 stars is plausible because the underlying content is real and the author has a strong distribution channel. A r04-… derivative at 4,000 stars is plausible only as a gaming artefact.

The single best heuristic before you fork a marketing skill repo is to check the commit history. Production-grade repos look like software (commits clustered around real changes, contributors over time, issues being closed). Derivative repos look like one big initial commit followed by README tweaks. The pattern is not subtle.

A second heuristic: look for tested_with:, version:, or validate-skills.sh in the repo. Three repos in the cluster ship a skill validator script (Eronred/aso-skills, manojbajaj95/claude-gtm-plugin, gtmagents/gtm-agents). Six repos use version: in skill frontmatter (thatrebeccarae/claude-marketing is the discipline reference). These are weak signals that the author treated skills as artefacts, not posts.

The cluster shape (small canon plus long tail) is what Step 5 of Lieben’s migration playbook produces at scale. The canon is operators with editorial discipline. The tail is everyone else.

The category map

Eight categories carry real depth. Two more carry surprisingly little. Counts below are repos with non-trivial marketing content (skills, scripts, frontmatter discipline), not raw search hits.

SEO is the deepest category

Five tiers, four canonical packs, dozens of derivative repos. The breakdown:

SEO is where the cluster is most mature. AgriciDaniel is canon for breadth; jdevalk/skills is canon for taste (the author was a co-founder of Yoast).

GEO and AEO are a real subcluster

Distinct from Google-rank SEO. The framing: tune content for being cited inside ChatGPT, Perplexity, Claude, and Gemini answers. The repos: zubair-trabzada/geo-seo-claude (6,978⭐), HouseofmvPS/claude-rank (170+ rules across ten scanners), mykpono/ultimate-seo-geo, and the seo-geo family of forks. The pattern is the same as the SEO tier above (rule-graph audits, schema markup, citability scoring) with the crawler set rotated to AI agents.

Email and outreach is shipping but missing the lifecycle layer

The marquee repos:

The gap inside email is lifecycle. The cluster has dozens of write a sequence skills. Almost none ship per-flow patterns (welcome, onboarding, re-engagement, win-back, abandoned-cart, churn-save, VIP, lapsed, referral) with copy plus send rules plus segmentation logic baked in. The strongest first-party Klaviyo and Braze coverage comes from thatrebeccarae, which happens to be the most under-distributed of the canon leaders.

Paid ads cluster around AgriciDaniel

The paid-ads sub-cluster is the second-deepest after SEO:

Paid ads is where the MCP cluster overlaps heavily with the skills cluster. Below.

Social media skews to LinkedIn and X

The strongest single-platform pack is sergebulaev/linkedin-skills (51⭐), with ten LinkedIn-specific skills (comment-drafter, engagement-monitor, hook-extractor, post-audit, profile-optimizer, reply-handler). blacktwist/social-media-skills (140⭐) covers ten cross-platform text-first skills. The hidden first-party pack is typefully/typefully, the only publish-grade skill from a real social-media SaaS, scheduling and publishing across X, LinkedIn, Threads, Bluesky, and Mastodon.

Brand and voice is conspicuously thin

This is the first of the four gaps the cluster carries. The canonical brand skills inside marketing repos are:

That’s four real brand-voice skill packs in a 480-repo cluster. None of them combine named voice profiles plus a forbidden-word lint plus rejection logic the way a hard brand system requires. clawfu-skills/branding/brand-voice-learner ingests existing content to learn voice; BrianRWagner/de-ai-ify removes AI tropes; neither commits to specific named voices nor enforces lint with rejection rules. The space is open.

Analytics goes through Google Search Console

Aminforou/mcp-gsc (779⭐), ahonn/mcp-server-gsc (214⭐), saurabhsharma2u/search-console-mcp (114⭐), Suganthan-Mohanadasan/Suganthans-GSC-MCP (57⭐), surendranb/google-search-console-mcp (30⭐), metehan777/google-search-console-mcp (21⭐), houtini-ai/better-search-console (12⭐), Magdoub/awesome-gsc-mcp (12⭐), serpfire/gsc-mcp-server (11⭐). Google Search Console is the most-MCPified marketing API by a wide margin. Nine separate MCP servers wrap it. The next densest API in the cluster (Meta Marketing) has three.

CRO is best-structured inside coreyhaines

The funnel-stage breakdown (page-cro, signup-flow-cro, onboarding-cro, paywall-upgrade-cro, popup-cro, form-cro) is the cleanest CRO taxonomy in any open marketing-skill pack. thatrebeccarae ships cro-auditor, ab-testing-framework, retention-churn-prevention, and landing-page-optimizer with ICE/PIE prioritization. JackChen-me/landing-page-doctor is a single-file 10-point first-screen scorer. The diagnostic side of CRO is well-covered. Experimentation skills sitting on top of real A/B platforms (Statsig, Optimizely, GrowthBook, PostHog) are mostly missing.

Industry-specific clusters are small but high-quality

Eronred/aso-skills (1,161⭐) is the App Store Optimization canon, fifty marketing-adjacent skills with a validate-skills.sh script. Weaverse/shopify-hydrogen-skills (33⭐) is the official Weaverse Shopify Hydrogen pack — first-party from a real CMS vendor. Affitor/affiliate-skills (374⭐) ships fifty affiliate skills covering research, content, infographics, landing pages, and deploy.

webflow/webflow-skills (64⭐) and jdevalk/skills (Joost de Valk) are the two strongest signals that established marketing vendors are starting to ship official skill packs. The marketing-skill cluster is shifting from operator-authored to vendor-authored over Q2 2026, and the Webflow and Yoast packs are the leading indicators.

Marketing-focused MCPs

The MCP cluster is denser on the analytics side than the campaign-execution side. The numbers:

The pattern: nine GSC servers plus three Meta servers plus fragmented coverage everywhere else. The MCP cluster is built where analysts already work (search console, GA4, Meta Ads Manager) and stays thin where campaign actions are riskiest (Klaviyo sends, Hootsuite publishes, TikTok bid changes).

The exception that proves the rule is kLOsk/adloop, which explicitly ships safety guardrails to prevent accidental spend. That’s the kind of policy layer the rest of the campaign-execution MCPs need before agents write to them autonomously, and it’s the kind of policy layer the playbook-style migrations underbuild.

The four gaps

The cluster is bigger and busier than any other marketing tooling space on GitHub. It still has four conspicuous holes.

Gap 1: brand-voice systems with hard lint rules

Plenty of brand skills generate copy. Almost none lint copy with rejection logic (forbidden words, no emoji in H1, no Get started CTA, must reference real numbers). The closest peers are blader/humanizer (removes AI tells), BrianRWagner/de-ai-ify (humanizer pattern), and clawfu-skills/branding/brand-voice-learner (ingests content to learn voice). None of them combine named voice profiles plus an explicit forbidden-word linter plus pass/fail tests.

This is the gap most under-served relative to operator demand. Every B2B SaaS team eventually writes its own banned-word list and copy-style rules. The cluster doesn’t have a canonical pack that scaffolds those rules into a Claude Code skill. The space is open for the first author to ship voice profiles for a dozen named B2B brands plus the lint scaffolding to enforce them.

Gap 2: lifecycle marketing as a flow library

Email skills exist in numbers. None of the canon leaders ship a per-flow library (welcome, onboarding, re-engagement, win-back, abandoned-cart, churn-save, VIP, birthday, lapsed, referral) with copy plus send rules plus segmentation logic. The flows live inside generic email-sequence skills that take what you’re trying to do as a free-text prompt rather than as a structured pattern. thatrebeccarae is closest with the Klaviyo and Braze first-party skills, but the per-flow pack hasn’t been built yet.

The reason this matters: lifecycle is the highest-yield email surface for SaaS and DTC, and it’s the surface where copy and rules are most coupled. A skill that generates an abandoned-cart email without knowing the cart-abandonment segmentation rule is just a draft. A skill that ships the segmentation, the trigger, the timing, and the copy together is a flow.

Gap 3: influencer, podcast, and affiliate execution

The cluster has plenty of strategy skills for each of these channels. It has very few operations skills.

For influencer marketing the only meaningful skill is clawfu-skills/social/influencer-finder, a single file. Discovery, outreach, contract, and reporting as a connected pipeline is unbuilt.

For podcast marketing, ericosiu/podcast-ops and clawfu/audio/podcast-* exist but lean production. The marketing side (booking, repurposing into clips and threads and newsletter, attribution back to website) is sparse.

For affiliate, Affitor/affiliate-skills is content and strategy. Affiliate operations (link tagging audits, leaderboard, payout reconciliation, fraud detection) is mostly missing.

The pattern across all three is the same: a strategy skill is cheap to write because it’s just a prompt with structure. An ops skill is expensive because it has to integrate with the affiliate platform’s API, the influencer’s contact graph, the podcast network’s calendar. The cluster is biased toward the cheap end.

Gap 4: marketing mix modelling and attribution

Yakoub-ai/agent-mmm is the only quantitative MMM skill in the entire cluster. Two stars. It uses pymc-marketing, the leading open-source MMM library. The fact that exactly one skill exists, and that it has not received meaningful adoption, is the cleanest evidence that the marketing-skill cluster is biased toward content generation and away from quantitative analysis.

There is no open-source skill for first-touch, last-touch, linear, or time-decay attribution audits. There is no skill that ingests a marketing budget and a list of channels and produces an attribution-aware spend recommendation. Operators run these analyses inside Looker, Mode, or BigQuery notebooks. None of the work has been packaged as a skill.

This gap is the most consequential for the long-run shape of the cluster. Marketing teams that adopt agent-driven workflows will eventually need their attribution analysis to live inside the same skill folder as their copy and ads. The team that ships that skill first will define the format.

Anthropic’s own marketing-shaped skills are deliberately broad

The marketing skills inside Anthropic’s official anthropics/skills repo (127K⭐) are broad rather than channel-specific:

There is no paid-ads, no email-sequence, no seo-audit inside the Anthropic repo. Channel-specific marketing skills come from the community. The official repo seeds the format and the design-side primitives. The marketing cluster fills in everything channel-shaped underneath.

This is the architectural choice that made the cluster possible. If Anthropic had shipped opinionated channel skills, the cluster would compete with the official repo and many fewer people would publish their own. Because the official repo is generic, every operator who runs a real campaign produces an artefact that doesn’t yet exist upstream. The skill folder is a contribution surface, not a competition surface.

The agency-as-distribution genre, full picture

Setting the marketing cluster next to the GTM cluster sharpens the size gap.

The marketing side is roughly 40× bigger. Two structural reasons: the universe of marketing operators is bigger than the universe of GTM engineers, and marketing channels are more numerous than sales channels. The marketing cluster also had a multi-month head start. The GTM cluster will likely close some of that gap by Q3 2026.

Both sides share failure modes (a small canon, a long tail star counts don’t separate, a bias toward content generation over operations and quantitative analysis). Both are downstream of Anthropic’s September 2025 decision to ship SKILL.md as a public format.

What to fork if you’re an operator

The cluster’s size makes choosing one repo harder than choosing zero. Concrete recommendations, ranked by what most operators actually need.

🔴 SaaS marketing — fork coreyhaines31/marketingskills. The reference template. Forty skills covering the full SaaS-marketing surface. The description triggers are the cleanest in the cluster. Don’t try to use all forty. Fork it, delete what you don’t need, and start from the funnel-stage CRO sub-skills (the strongest pieces).

🔴 SEO — fork AgriciDaniel/claude-seo. Twenty-four sub-skills plus twelve specialist subagents. If your motion is content-led, this is the depth reference. Pair it with zubair-trabzada/geo-seo-claude if you want AI-search visibility.

🔴 Paid ads — fork AgriciDaniel/claude-ads. Two hundred fifty checks across six platforms. Pair with kLOsk/adloop for the spend-safety guardrails.

🟡 Versioning and tested fields — fork thatrebeccarae/claude-marketing. Twenty-six stars, fifty-six versioned skills with tested_with: metadata. The most production-disciplined repo in the cluster. The first-party Klaviyo and Braze skills are the only ones that exist at this quality.

🟡 GTM-specific — see the playbook section of the ColdIQ audit. Fork gtmagents/gtm-agents for B2B sales-led teams or growthenginenowoslawski/coldoutboundskills for the cold-outbound flow.

🟡 Brand voice — wait for the gap to fill, or build it. No canonical pack exists. The closest peers (blader/humanizer, BrianRWagner/de-ai-ify, clawfu/brand-voice-learner) each handle one leg of the brand-voice problem. The first repo to ship named voice profiles plus a forbidden-word lint plus rejection logic will own this category.

🟢 Resist forking everything. The long tail is mostly noise. Star counts are not a quality signal inside this cluster. Commit history is. Three or four well-chosen repos forked once will out-perform forty repos forked optimistically.

🟢 Skip the cluster entirely if your motion doesn’t match. Operators with sales cycles measured in days, with non-digital channels, or with regulatory environments that won’t approve agent actions on customer data, get less rent from this work than the migration cost. The marketing-skill cluster pays back on volume and asynchrony. Without both, the setup overhead is bigger than the win.

What’s worth bookmarking

Three pointers to keep within reach if you’re navigating this cluster more than once.

🔴 anthropics/skills (127K⭐) — the format spec and the design-side primitives (brand-guidelines, theme-factory, frontend-design, web-artifacts-builder). Read the README and the existing skills before you write your first one. The format conventions are load-bearing.

🔴 ComposioHQ/awesome-claude-skills and VoltAgent/awesome-agent-skills — the two awesome-lists with real curation. Most other awesome-lists in the cluster are programmatic forks; these two are hand-maintained and have an actual Marketing/Business section.

🔴 coreyhaines31/marketingskills (26K⭐) — the canonical SaaS-marketing template. Even if you don’t fork it, read three of its skill files (copywriting, cold-email, page-cro) before writing your own. The trigger language and the funnel-stage breakdown are the closest the cluster has to a style guide.

The cluster will keep growing through Q3. The canon will probably stabilise around the four leaders named above, with vendor-authored packs from Webflow, Yoast, Typefully, and others filling in the channel-specific tiers. The four gaps (brand-voice lint, lifecycle flows, influencer-podcast-affiliate ops, MMM and attribution) will each get filled by an operator who does the work first and open-sources it. Step 5 of the playbook is the engine that produces those repos. The cluster is the proof that the engine works.

Not every operator open-sources. The personal-brand variant of the same flywheel runs gated: agency publishes the system architecture as an X long-form, withholds the actual skill files, converts inbound through sales calls instead of repo stars. See WizOfEcom’s 5-agent content engine, audited for the gated counterpart to the open repos covered here.