I Reviewed Corey Haines' Marketing Skills So You Don't Have To. There's Good, Bad, and Ugly

Corey Haines' marketing skills have 4.6 million installs. 13 of the 50 beat playbooks you'd pay for. Most of the rest are fine once you skip the pep talk. 4 will waste your time or make you wrong. I read all 260,000 words so you don't have to.

The descriptions are keyword-stuffed to trigger on as many prompts as possible. 44 of 50 use the same “also use when the user mentions” construction, and one commit reads “optimize all 32 skill descriptions for better triggering.” And it works: installs run from 199,300 at the top to 15 at the bottom. Fair play. The problem is what happens next. A skill steers your agent with far more authority than anything you ask it to research, so a bad one costs more than a bad blog post.

I graded the standouts and stragglers against subject matter experts, not Google’s documentation: Ronny Kohavi on experimentation, Peep Laja on conversion optimisation.

Bad vs ugly

Bad is the CRO skill telling you to check for white space. Ugly is the schema skill telling you FAQ markup wins rich results. You only catch ugly by checking, and nobody installing a skill checks.

The Good

attribution · ranked 68/69 with 15,400 installs and 9,135 words

Take it.

The best thing in the pack by a distance. It opens with what it does not own and routes you to 4 other skills. It sets expectations before touching a number:

Attribution is directional, not truth… Every model is an opinion. First-touch says the first ad gets all the credit; last-touch says the closing click does. Both are wrong in opposite directions.

Every blog has the table of 6 attribution models. This one adds a column no blog has: how it lies. Last-touch “over-credits bottom-funnel and branded search; ignores what created demand.” Data-driven: “a black box; needs volume; can’t see offline or dark touches it was never fed.” That column is the judgment you’d otherwise pay for.

It also covers the methods that establish causality, which is where the field went after tracking collapsed. Mentions in the file:

Ghost ads are the tell. Most working marketers have never heard of them.

It handles a perishable fact correctly too. It states Google’s old data-driven attribution thresholds, notes Google has since dropped them and made DDA the default, and keeps the judgment that still holds: “low volume still makes it noise dressed as science.” I checked. All 3 are right. Another skill in this pack is still teaching a feature Google killed in 2023.

Use it before any budget reallocation. Its core warning: multi-touch attribution is directionally biased. It over-credits retargeting and branded search and dumps what it can’t follow into direct.

offers, marketing-plan, ad-creative and prospecting are the same quality. All are only as good as the model you run them through, so sanity-check the occasional terrible suggestion.

The Bad

cro · ranked 25/69 with 60,900 installs and 3,820 words

Leave it.

Conversion optimisation is the first discipline in the repo’s own description. The file doing the work is 897 words and has never had a commit since the one that created it.

Its whole framework is 7 headings “in order of impact”: value proposition, headline, CTA placement, visual hierarchy, trust signals, objection handling, friction. That claims the same order on every page, in every market. Peep Laja, whose CXL is the most-cited CRO training:

Amateurs follow best practices and don’t know where to begin. Experts follow frameworks and processes.

Laja’s ResearchXL runs 6 research stages before anyone proposes a change. This skill has no research stage. Session recordings, heatmaps and user research appear once, as question 4 of a closing questionnaire printed after the recommendations.

Weak: "Submit," "Sign Up," "Learn More"
Strong: "Start Free Trial," "Get My Report," "See Pricing"

Is there enough white space?

None of it is wrong. It’s every conversion listicle since 2012, which the model produces unprompted. The cost: your agent hands you a confident page critique that never looked at a session recording, and you act on it, lured by the trust his good skills earned.

ab-testing · ranked 45/69 with 53,200 installs and 3,972 words

Better. The sample-size table is correct, it warns against peeking, it mandates guardrail metrics. But check it against Kohavi, who ran experimentation at Amazon and Microsoft:

Kohavi calls an experiment without a sample-ratio-mismatch check a car without seatbelts. A split of 50.2% instead of 50.0% is enough to stop the analysis. 4,000 words on A/B testing never mention it.

It also contradicts itself. It targets 4 to 8 experiments a month. Its own table needs 7,000 to 150,000 users per variant. Laja: “most SMBs don’t have the transaction volumes to run even a single test per month.” The skill takes your traffic number, then designs a test regardless.

Verdict: leave cro. Use ab-testing only once you’ve confirmed you clear roughly 500 conversions a month.

What to use instead

I wrote cro-research. No checklist. It reads your GA4 exports, flags which pages and devices convert worse than the rest of your site and where the funnel leaks, and marks every flag NEEDS AN ANSWER, because a number says where, never why. It also does the 2 things ab-testing skips: works out from your own traffic whether you can test at all, and checks the visitor split before showing a result.

npx skills add BigChrisCooke/seo-playbook-skills --skill cro-research

Not mine: ab-test-analysis from nimrodfisher runs the same split check in plain Python.

For marketing-psychology, install nothing. Your model already knows all 72.

The Ugly

schema · ranked 40/69 with 54,500 installs and 1,543 words ai-seo/content-patterns · ranked 8/69 with 118,300 installs and 1,435 words

Delete these.

Both produce output that’s factually wrong, confidently.

1. schema teaches 2 dead Google features

Its table still lists FAQPage and HowTo as routes to rich results. Google removed HowTo rich results in 2023 and deprecated FAQ rich results on 7 May 2026. The table was written on 26 January 2026 and never changed, so HowTo was 2 years out of date the day it was typed.

Add the old schema-markup listing’s 56,100 installs and this is 110,600, which would rank 9th.

Worse, the skill’s own test file grades an agent correct for recommending it:

"prompt": "…how do I add FAQ schema to get the rich
   results in Google?"

"assertions": [
  "Explains rich result benefits",
  …
]

The right 2026 answer, keep the markup for AI answer engines and expect nothing from Google, fails this test. The evals enforce the rot.

2. A pattern that says “always include sources” demonstrates with a fake one

Inside the otherwise-decent ai-seo skill is a reference file of patterns for content that gets cited by AI assistants. The first pattern:

Statistics increase AI citation rates by 15-30%.
Always include sources.

**Example:**
According to Google's 2024 Core Web Vitals report, 70% of
web traffic now comes from mobile devices, and pages failing
mobile usability standards see 24% higher bounce rates.

There is no such report. I looked. The worked example that teaches sourcing is a fabricated citation with a named institution, a year and 2 confident percentages. Follow it at scale and you publish content that looks rigorously sourced and isn’t. The same file quotes a real 115% citation-lift figure from the Princeton and IIT Delhi paper at KDD 2024, and never names it. To the agent, the invented stat and the real one are indistinguishable.

Fix: strike the 2 rows from the schema table and remove the Statistic Citation Block. 90 seconds. The rest of ai-seo is worth keeping.

Or install the corrected versions

His repo is MIT-licensed, so I published corrected copies. Each keeps his copyright notice and carries a dated comparison against his version.

• schema-markup-better-than-corey-haines separates “valid schema.org” from “eligible for a Google rich result”, and drops FAQPage, HowTo and the sitelinks search box as routes to one.
• ai-seo-better-than-corey-haines replaces the invented citation with checked primary sources and ships a scanner that flags undated claims.

npx skills add BigChrisCooke/seo-playbook-skills --skill schema-markup-better-than-corey-haines ai-seo-better-than-corey-haines

Not mine: seo-schema in AgriciDaniel’s claude-seo carries a dated status table with FAQ retired on 7 May 2026 and HowTo removed in September 2023.

The names claim those specific fixes, not that the whole pack is better. Check them the way I checked his.

The filter

Small numbers are each skill’s install rank out of the 69 listed.

Keep product-marketing despite its thinness. The other 49 skills read its output file.

The install count measures the wrong thing

The skills worth taking sit, on median, 60th by installs. The ones to avoid sit 25th.

The worst skill has 9 times the installs of the best. Installs track age and keyword surface, and skills.sh has no reviews or ratings, so the only number in the ecosystem measures discoverability.

And 25th undercounts cro. The May release folded page-cro and form-cro into it, and the old listings still collect installs.

Around 380,000 agents have been handed that checklist as their model of conversion optimisation. The worst thing in the repo sits at the top of its most popular discipline.

What separates good from bad

Not AI. 320 of the 469 commits are AI-co-authored, including attribution and offers, the 2 best things in the pack. The variable is whether someone who knows the discipline was steering. attribution reads like a practitioner using a model to write faster. cro reads like nobody was in the room, so the model returned the average of everything ever written about conversion optimisation. Unsteered, a model gives you the mean of its training data. In marketing, the mean is a 2012 blog post.

I ran the same lint over 4 other packs. The ones with something to sell, Squirrelscan’s CLI and Hyper’s MCP, run far more steps and cite far more sources than the free advice packs. They have to work in front of a customer.

None of this makes Corey Haines a villain. He built something people want, gave it away, and kept the good parts genuinely good. 13 strong skills, a long middle that works once you skip the pep talk, and 4 files that shouldn’t be in anyone’s context window.

Just don’t let the install count do your thinking. It measures distribution, not whether the advice is any good.