Canadian marketer from UA LLM | SEO | Growth Hacking
Canadian marketer from UA LLM | SEO | Growth Hacking
Field report · Agent skills
A practical review from an agency SEO lead — what we kept, what we dropped, and what it actually costs.
SEO operations · 9-person agency · 8 min read
When Anthropic shipped Agent Skills, my first reaction was fatigue. Another packaging format for prompts. We'd already been through custom GPTs, prompt libraries in Notion, and two internal "AI playbooks" nobody opened twice. So this isn't a hype post. This is an inventory of what's still installed on our team's machines six months later, after the novelty wore off and the only filter left was "does this save billable hours."
Quick context so you can calibrate against your own situation: we're a nine-person agency, mostly mid-market B2B clients, retainers built around content and technical SEO. Reporting eats roughly 20% of our month. That number is the villain of this story.
A skill is a folder with a SKILL.md file — instructions, workflow, sometimes reference docs and templates. Claude loads the short description at startup and pulls in the full instructions only when a task matches. That's it. No fine-tuning, no vector database, no platform lock-in. The format is boring in the best way: it's markdown in a git repo, which means you can read exactly what the skill will do before you run it, diff versions, and fork anything that doesn't fit your process.
That auditability turned out to matter more than I expected. Half the AI tools we've trialed are black boxes. Skills are the opposite — when a deliverable comes out wrong, you open the file and see which step produced the garbage.
The single biggest workflow change came from SE Ranking SEO skills — a catalogue of twenty production skills built on top of their MCP server, so the outputs are grounded in live ranking, backlink, and SERP data rather than the model's memory. That distinction is everything. A content brief hallucinated from training data is a liability; a content brief assembled from actual keyword metrics and a real competitor teardown is a deliverable.
Three of them are now part of our standard retainer workflow:
seo-content-brief replaced a process that used to take a strategist half a day. It pulls keyword research, analyzes the live SERP, tears down competing content, and proposes internal links. Our editors still rewrite the angle — the machine picks safe angles — but the research layer is done.seo-drift is the one I'd defend in a fight. It baselines a site (authority, traffic, keywords, backlinks, page fingerprints) and diffs against that baseline later, with severity-coded regressions. We run it after every client deployment. It has caught two botched migrations before the client noticed.seo-competitor-gap-analysis feeds our quarterly planning. Keywords competitors rank for that the client doesn't, filtered by intent and difficulty. Nothing conceptually new — every SEO tool has a gap report — but here the output arrives as a prioritized plan, not a 4,000-row export someone has to interpret.Honest caveat
The Data API calls charge credits, and the big skills (full gap analysis on a large domain, keyword clustering with many seeds) can burn through thousands of them. Check your balance before pointing it at an enterprise site, and use the cost ceilings the skills document. We learned this the expensive way.
AgriciDaniel/claude-seo takes a different architectural bet: instead of one skill per deliverable, it's a universal orchestrator with 25 sub-skills and 18 agents it dispatches in parallel. Point it at a site, and it detects the industry, fans out specialists, and synthesizes an action plan where each recommendation carries the evidence it rests on and a "how would we know this failed" check. We use it for new-client audits — the falsifiability framing alone has improved how our junior analysts write recommendations.
Not strictly SEO, but coreyhaines31/marketingskills covers the adjacent work every SEO retainer bleeds into: page CRO, copywriting, comparison pages. The clever bit is a shared product-marketing context file every skill reads first, so outputs stay consistent with the client's positioning instead of resetting to generic each session.
I won't name them, because they're free and someone worked hard on them — but the outputs were confident, structured, and unverifiable. That's the actual risk profile of this ecosystem: the format makes everything look production-grade. The filter we settled on — where does the data come from? If the answer is "the model," treat the output as a draft. If it's an API or a crawl, treat it as a report you still spot-check.
The onboarding curve deserves a sentence: the first two weeks, people trusted outputs too much, then too little, before landing somewhere sane. Budget for that adjustment the way you'd budget for the setup itself.
Would I rebuild the workflow around skills if we started today? Yes — but with the spot-checking discipline in place from day one, rather than learned through embarrassment.
Canadian marketer from UA LLM | SEO | Growth Hacking
Canadian marketer from UA LLM | SEO | Growth Hacking