How to Use Corey Haines' Marketing Skills for Claude Code

The marketingskills repo went from launch to over 41,000 stars in six months, and almost everything written about it is a ranked list that mentions it in one line. This is the part those lists skip: getting it installed where your agent can actually see it, and knowing what each skill will and won't do once it's there.

What it actually is

It's a collection of 47 marketing skills, each one a plain Markdown SKILL.md file under skills/<skill-name>/. No build step, no runtime, no service. Skills cover CRO, copywriting, SEO, ads, analytics, email, and growth engineering. Many ship with references/ for deeper material the agent loads on demand, and 45 of the 47 carry an evals/evals.json.

If you're still deciding whether to run marketing from an agent at all, that case is made here. This guide assumes you've decided and want the repo working.

Small thing worth knowing before you quote a number: the README says "45+", the plugin marketplace manifest says 47, and the tree has 47. Count from the tree.

Install it into the folder your agent reads

The recommended path:

npx skills add coreyhaines31/marketingskills

This is where most people lose an hour. If you ask your agent to run that command for you rather than running it in your own terminal, the CLI runs non-interactively and may install only into the universal .agents/skills/ directory. Claude Code does not read that directory. The repo's own README flags this. Pass the agent explicitly:

npx skills add coreyhaines31/marketingskills -a claude-code

If you prefer the plugin route, inside Claude Code:

/plugin marketplace add coreyhaines31/marketingskills
/plugin install marketing-skills

There's also git clone + copy, a submodule setup, and npx skillkit install for multi-agent environments. You can also install a subset with --skill cro copywriting, which is worth doing since a full install drops 47 skills straight into your skills folder with no namespace prefix, mixed in with whatever you already had. There's an open request for an mkt-* prefix; it isn't implemented.

Run product-marketing first

This is the step that changes how everything else behaves. product-marketing is the foundation skill: it writes .agents/product-marketing.md, and every other skill checks for that file before doing anything. It can auto-draft a first version by reading your README, landing page, and marketing copy, then asks you what to correct.

Skip it and every skill re-interrogates you about your product, audience, and positioning. Run it once and the rest inherit the context.

What ad-creative actually does

The headline skill runs in four modes:

1. Generate from scratch using your product context, audience, and platform conventions.
2. Iterate from performance data. You hand it CSV or pasted results; it analyses winners and losers, then generates variations that extend what worked and test one or two new angles.
3. Scaled static batches, on the order of 50 concepts per run.
4. Creative strategy loop, which decides which ads are worth making at all by combining account performance, customer language, and external organic signal into a ranked roadmap.

Mode 3 has a hard prerequisite that I'd call the most honest thing in the repo. The skill states plainly that most AI ad generation fails on input grounding, not output quality, and it enforces that: if inputs/winning-ads/ or inputs/reviews/ is empty, it stops and asks you to fill them instead of generating ungrounded concepts. It also refuses to invent claims, stats, or testimonials.

The practical consequence: a brand new product with no ad history and no reviews cannot use the scaled mode. That's a real constraint, and the repo is upfront about it.

The tools directory doesn't ship with the skills

There are 64 zero-dependency Node CLIs under tools/clis/ for pulling live data from Ahrefs, Google Ads, and similar. They read credentials from environment variables and need Node 18+ with no npm install.

But the plugin manifest points only at ./skills, so npx skills add and the plugin marketplace both give you the skills without the tools. To use them you have to clone the repo. The README doesn't spell this out; it falls out of the manifest.

If you're upgrading from v1.x

v2.0 renamed 17 skills and merged page-cro and form-cro into cro. New skills install alongside the old ones rather than replacing them, so you'll end up with both until you delete the old folders by hand. The README carries the exact rm -rf list. Move your context file too:

mkdir -p .agents
mv .claude/product-marketing.md .agents/product-marketing.md 2>/dev/null

Skills still fall back to the old locations, so this one is optional.

Honest limitations

• No install path for claude.ai web or desktop users. All 22 releases ship zero assets, so there's no .skill bundle. The maintainers acknowledge this in an open issue, describing non-technical marketers on claude.ai as arguably the largest unserved segment.
• Releases lag main. The latest tagged release is v2.6.0 while the plugin manifest is already on 2.8.12. Don't judge freshness by the Releases page.
• A security scanner may halt your install. One user reported skillkit aborting partway through on a HIGH severity finding against competitor-profiling, triggered by a line instructing the agent to proceed without asking.
• Codex support is community-improvised. The README claims Codex compatibility and the repo is tagged for it, but there's no Codex plugin manifest; the open issue's workaround is dropping skills into ~/.agents/skills.
• The repo is a funnel. The README links the author's paid products at the top with ?ref=marketingskills attribution. The content is genuinely MIT and genuinely useful, and the commercial motive is visible rather than hidden. Worth knowing, not a reason to avoid it.

The grounding problem, and where it points

That line about input grounding is the most transferable idea in the whole repo, and it generalises past ads.

Ask an AI for an ad with nothing but a product description and you get plausible, generic output. Give it 20 ads that actually converted and the same model produces something worth testing. The constraint was never the generator. It was what you fed it.

Short-form video has exactly this problem, and a harder version of it, because you can't grade a video the way you grade a headline. This is what Riffkit is built around: instead of describing what you want, you give it one video that already worked, and it rebuilds that formula, meaning the hook, the pacing, and the emotional beats, as new footage with your product in it. It never re-uploads the source clip. Output is post-ready, generated natively in 9 languages.

The grounding difference is worth being precise about. ad-creative's scaled mode needs your own 90 days of winning ads, which a new product doesn't have. Riffkit needs one proven video, and it doesn't have to be yours. That's the whole point: the formula that worked for someone else in your category is evidence you can use on day one.

Riffkit's own video skill installs the same way these do, so the agent you just set up can run it:

Read https://riffkit.ai/SKILL.md and follow the instructions to set up Riffkit.

Then ask in plain language: "riff this TikTok into a video for my product." If you're deciding between generative video tools, the MoneyPrinterTurbo alternatives breakdown covers where each approach fits.

And go star the repo. 47 well-structured marketing skills under MIT is a serious contribution, whatever else it's upstream of.