Marketing OS: I Let a Claude Skill Audit My Site
An open-source Claude skill claims to be a marketing department in one file. I ran its audit on my own site, kept the score, and checked every finding.
There’s a repository on GitHub called Marketing OS that describes itself as “a complete marketing department in one skill.” It has 472 stars, an MIT license, and not a single line of executable code - the whole thing is markdown. Fourteen instruction files that tell an AI agent how to audit a website, score it out of 100, write copy, diagnose ad fatigue, and a dozen other jobs a marketing team does.
Reviewing a thing like this from its README would be malpractice, so I did the obvious experiment instead: I pointed it at my own site. This site. The one whose whole pitch is that it practices what it publishes. Marketing OS gave aiforthemarketer.com a 66 out of 100, and the most useful thing I can do with that number is show you exactly how much of it survived checking.
What it actually is
Marketing OS is an agent skill: a set of instructions an AI assistant loads when a task matches. The architecture is a small router file that reads first, plus fourteen reference modules that load only when needed - audits, GEO, copywriting, hooks, paid ads, email, social, launches, positioning, competitor teardowns, app stores, analytics, and a catalogue of AI writing tells it runs over its own output before delivering anything.
Three rules run through every module, and they’re the reason I took it seriously enough to test. Score everything, because findings without a number are easy to argue with. Ship artifacts, not advice - write the replacement headline rather than describing it. And state what you couldn’t determine, in an explicit gaps section, in every report. The honesty rules go further than most human consultants do in writing: never invent proof, never declare winners on small samples, all scores are labeled as heuristics, and never anchor on the number the client suggests.
If you use Claude Code, installation is two commands (/plugin marketplace add Yuzzyuk/marketing-os, then /plugin install). In the Claude apps it’s a zip upload under Settings - and it sits naturally alongside Claude Projects, which hold your standing brand context while a skill like this instructs behavior. I ran it slightly differently for this review - I loaded its instruction files and followed them exactly as written - which also means I read all fourteen modules in full, something I’d recommend to anyone before letting an agent skill loose on their business.
How the audit ran
The audit module turns one URL into a scored report across six weighted dimensions: messaging (25%), conversion (20%), search (20%), competitive position (15%), trust (10%), and growth (10%). Its instructions say to fan the six dimensions out to parallel subagents, each carrying its slice of the scoring rubric, then synthesise the total yourself - with the pointed rule that the synthesis is never delegated, because “the pattern across dimensions is the product.”
So that’s what happened: six agents, each fetching my live pages against its rubric, running in parallel. The fan-out finished in under three minutes. The whole exercise, synthesis and my own verification included, took about half an hour. A human consultant delivering the same document shape - scorecard, prioritised fixes with replacement copy written out, an explicit list of what couldn’t be inspected - would take days and charge accordingly, and that comparison is the entire commercial argument for skills like this. Whether the document is right is a separate question, which is where this gets interesting.
The scorecard it handed me
| Dimension | Score | Its one-line verdict |
|---|---|---|
| Messaging & positioning | 79 | Strong. The independent “this site sells nothing” stance is the hardest thing here for a competitor to copy, but the homepage headline never says what the reader gets |
| Conversion | 48 | Actively leaking: no email capture anywhere on the site |
| Search & discoverability | 68 | Solid structure, but flagged “missing” meta descriptions and schema |
| Competitive position | 70 | Only independent voice in a space of vendor content, and says so too quietly |
| Trust & credibility | 68 | Excellent citation discipline in posts, thin trust signals around them |
| Growth & retention | 58 | Good architecture, zero lifecycle hooks - a reader can’t choose to come back |
Weighted total: 66/100. The rubric’s own calibration note says the median real site lands between 55 and 70, so this is “functional but generic” territory - a phrase I’ve now had the character-building experience of reading about my own work.
The synthesis line it produced was genuinely good: every low score traces to the same root, a site that publishes rigorously and captures nothing. No newsletter, no social presence, no way for a reader who liked a post to ever hear from me again. That’s not a subtle observation, but it’s the correct one, and having it arrive with a number and a priority order is what turned it from a thing I vaguely knew into a thing I’m fixing.
What it got right
Four findings survived verification, and I’m acting on all of them.
The email capture gap is real, and scoring it as the top conversion problem rather than a growth nice-to-have is the right severity. The content earns attention, and the site has no way to keep it.
The sourcing policy is invisible. Every statistic on this site links to its original source - that’s the house rule and the differentiator - and the audit correctly noticed you can’t learn that from the homepage. It sits in paragraph nine of individual posts. “Your defensibility is your least visible asset” is the kind of finding worth 30 minutes of anyone’s time.
The competitor read was sharp. The competitive agent fetched three sites ranking for the same topics and found all three were guides functioning as lead generation for a platform, an agency or a consultancy. Its conclusion, that independence is the rare position here and should be shouted rather than mentioned, matches what I believed but had never tested against actual competitor pages.
The title-length catch was one I’d missed. My build enforces 60-character headlines, but the rendered page appends the site name, pushing long titles past what search results display. A small true thing no checklist of mine had caught.
Where it went wrong, and why that’s the real lesson
Now the other half. Two of the three search findings were false positives. The audit reported my meta descriptions and my structured data as missing. Both exist on every page - the build literally fails if a description is absent - but the fetch tool the agents used converts pages to readable text and strips the HTML head where both live. One agent even listed “schema may exist but be stripped by my tools” in its own unknowns section, then scored the signal as a FAIL anyway.
That move has a name in the skill’s own rubric. It instructs agents to use four verdicts - pass, fail, unknown, not applicable - and states plainly that converting unknowns to fails “inflates urgency” and is one of the two lies a scorer can tell. The instruction was right there. The model broke it anyway, because models resolve uncertainty toward whatever produces a complete-looking answer.
Same pattern, second offence: the honesty spine says never invent proof, no fabricated statistics, not even as placeholders. The audit’s findings respected that. Its rationales didn’t - justifications like “23% higher CTR on average” and “proven 10-25% referral lift” appeared with no source attached, exactly the confident filler the skill’s own slop-catalogue exists to catch. And a passage in one of my posts that is explicitly labeled as an invented illustration got flagged as an unsourced statistic, which is what happens when an agent pattern-matches “number without citation” without reading the sentence above it.
None of this makes the skill bad. It makes the skill honest about something its README can’t fix: markdown can instruct a model, but it can’t make a model obey. The rules are genuinely excellent. The compliance is probabilistic. Which means the score’s job is to direct your attention, not to be the verdict - my 66 became a 68 after an hour of checking, and the checking is not optional.
The GEO module, judged on the subject I know best
AI-search citability is the topic this site covers most deeply, so I read that module hardest, looking forward to catching it saying something stupid. It doesn’t. It refuses to promise citations, tells you no engine publishes its criteria, insists a single query observation is noise, and warns against exactly the over-claiming that most GEO advice is made of. Its llms.txt section - adoption marginal, effect unestablished, recommend it as cheap hygiene and never as a lever - is more honest than most paid consulting on the subject, and its warning that Google retired HowTo and FAQ rich results stops a recommendation that half the industry is still copy-pasting from 2023.
I ran its five citability levers against this site and got the most useful sentence of the whole exercise: my binding constraint is corroboration. The on-page work scores near its ceiling - that’s the house obsession - but almost nobody else on the web talks about this site yet, and models weight what others say about you above what you say about yourself. No on-page edit fixes that. It’s a link-earning problem, and I’d been avoiding that conclusion because the on-page work is the part I enjoy.
Verdict
Run it. It’s free, it’s readable in an hour, and the fourteen files are worth reading even if you never execute them - the audit rubric and the honesty rules are better editorial standards than most marketing teams have written down anywhere. My run produced four real fixes, one uncomfortable strategic truth, and a scorecard I mostly trust because I checked it.
But treat every FAIL as a lead, not a verdict. The failure modes I hit - tool artifacts scored as site defects, unknowns hardened into fails, unsourced numbers decorating the reasoning - will be in your report too, wearing the same confident prose as the true findings. The skill’s own three rules tell you what the deliverable should be. What they can’t tell you is which lines the model actually followed them on. That part is still your job, the same freelancer test I apply to every AI draft: accept the work the way you’d accept it from a talented junior with a superb methodology binder and no fear of being wrong.
One number to hold onto: 66. That’s what a well-designed heuristic said about a site I’ve spent months making rigorous, and about half the gap it found was real. In a field where every tool promises transformation, a free markdown file that’s half right, checks in thirty minutes and admits what it couldn’t see is - and I mean this as high praise - the most honest marketing product I’ve tested this year. More tool tests are coming in the tools category, and the practice they plug into lives in the complete guide to AI marketing.