Marketing skills for Claude: what a good one looks like, and where skills stop

A skill is a folder. Its SKILL.md needs two fields: a name and a description. Claude loads in three steps:

1. The name and description of every skill, all the time. About 100 tokens each.
2. The body of SKILL.md, only when your request matches the description.
3. Reference files and scripts, only when the body points to them. A script runs and only its output comes back.

That is why you can install many skills without slowing Claude down. It is also why the description matters more than anything else in the file.

The description is what Claude matches your request against. “Google Ads expert” fires on nothing. Say what it does and when to use it, in the words you would type:

Anthropic suggests keeping the body under about 500 lines. Ours hold the method: what to check first, in what order, what counts as evidence, and the shape of the answer. The depth goes in reference files the body names:

• A metric dictionary. What each number means, whether it adds up, and what to watch for. Reach and frequency do not add up across ads. We once saw an account read 66,189 reach that way; the true number was 36,933.
• Diagnostic patterns. A signal, its causes in order, and the one read that tells them apart. Our Google Ads skill has 20.
• How each platform takes credit. Meta, Google Ads, Analytics and your store will each claim the same sale. Without this, Claude adds them up.

This one is easy to miss. When Claude looks for the part of a reference that answers you, a heading like “High click-through rate, low conversion rate” wins. A heading like “Section 4.2” loses to whatever the file’s introduction says. We watched the wrong passage win exactly that way, so every reference we write now has one idea per heading.

Instructions ask Claude to do something. A script proves it did. Our skills carry tools that check the output: a linter that blocks copy with the tells of machine writing, and a checker that reads the built site to confirm every published page holds the article it claims to. Claude runs them and only the result comes back.

A good marketing skill says what it cannot do. Ours say it plainly: read-only, never spends, never publishes. That one line stops Claude from offering to “just go ahead and pause those ads” with no approval behind it.

Even a great skill is know-how with no hands. It cannot:

• see your accounts unless you paste the data in;
• remember what happened on a client last month;
• make a change and check that it took;
• tell you next month whether its advice worked;
• notice a problem while nobody is asking.

In our own accounts the costly problems were exactly those: an Analytics account that could not name the campaign behind 107 paid visits, a change Google applied that nobody approved, a test that never ran. No skill finds those. Connected data does.

What we built on top of the skills