On 26 October 2026 we ran three skills from a GitHub repo that has been circulating widely against our own website, then checked every finding against the live pages. The most useful result was an uncomfortable one. Our home page sends the main button straight to a €990 checkout without saying what the Marketing Review contains. On our AI agent orchestration page, three different engagements (Discovery Sprint at 2 to 3 weeks, Prototype Build at 4 to 8 weeks, Agent Operating System at 8 to 16 weeks or more) all link to that same €990 checkout, and the price itself never appears on the page. All three skills found some version of this offer problem.
What the repo is
The repo is github.com/coreyhaines31/marketingskills by Corey Haines, released under the MIT license. It is a library of Agent Skills for coding agents such as Claude Code, Codex and Cursor. At the commit we tested (e85f3cb, 19 Oct 2026) it contained 52 skill folders. Every skill first reads a "product-marketing" context file that describes your product. GitHub showed 56,578 stars on 26 Oct 2026 at 09:47 CET. In the week of 20 October several posts on X promoted the repo to large audiences, and one by @RoundtableSpace had about 65,800 views.
The README says "Verified Partners" fund the work and are "listed alongside the neutral options and never influencing what the core skills recommend." Keep that in mind when a skill recommends a specific tool.
How we ran it
We ran cro, seo-audit and copywriting against two live pages: our home page and our AI agent orchestration page. Each skill's full instructions were the system prompt, and the page data came from a live fetch of the HTML. The model was Claude Sonnet 5.5 via OpenRouter. We had no product-marketing context file, so we told the model to state its assumptions and not to invent numbers. We changed nothing on the site. Afterwards we checked every finding against the live pages with our own fetches.
We checked 22 findings. 14 held up, some only partly or as a fair opinion. 3 were wrong or overstated, 3 were weak or judgment calls, 1 was a cluster of generic advice, and 1 was caused by our own setup.
What held up
Besides the offer problem, the skills found smaller issues that are easy to stop seeing on your own site. Two of our "where growth gets stuck" cards read like solutions instead of problems. The "Guides" and "Cases" links in the navigation go to the same page. The footer still says "Updated: August 2026". Our thank-you pages carry a noindex tag, but robots.txt blocks those pages, so Google never sees the tag.
The copywriting skill was the most careful of the three. It wrote a new H1 for the orchestration page, "AI agent workflows your team can check, approve and audit", using only claims that were already on the page. It also listed what it would not write without facts: delivery time and refund policy. We found that more useful than a confident guess.
What was wrong, and what was our fault
All three skills flagged an image path with "imges" in the folder name as a likely typo that might be broken. The image loads fine. The CRO skill said our Contact page might return a 404. It loads, and the lowercase URL redirects to it. The SEO skill estimated our meta description lengths a few characters off; we measured 157 and 150 characters. Every "may be broken" claim was a guess, and a single fetch settled each one.
One finding came from our own setup. The SEO skill said our sitemap had no language alternates. It has 122. We had only passed it the list of URLs, not the full file. The same run saw only the first 1,500 characters of our structured data because our extraction script cut it off, and the skill pointed that out itself. The skills can only judge what you give them.
The generic cluster was the usual list: add a guarantee, add a testimonial, run these A/B tests. These suggestions would fit almost any page, so they told us nothing specific about ours.
If you want to repeat this, one technical detail: our first SEO run returned nothing because the model spent its entire token budget on reasoning. We ran it again with a larger budget and got a full report.
How we would use these skills
If we run them again, we will follow four rules:
- Fill in the product-marketing context file first. Without it, the model has to guess your positioning.
- Pass complete page data: the full sitemap, the full structured data and the full HTML. With cut-off input, the findings are incomplete but sound just as confident.
- Check every "may be broken" claim with a real fetch. Our wrong findings all sounded plausible, and each took seconds to check.
- Treat the output as a checklist for a person who decides (we wrote about that split in "Agents do the busywork. People make the call."). 8 of our 22 findings needed a person to downgrade, discard or reclassify them.
We have not fixed the issues yet. The checkout links, the outdated footer, the duplicate navigation entry and the rest will go through our normal release process like any other change. The skills found real problems on our pages. Deciding what to change, and shipping it, stays with us.