Claude skills for marketing: a full example, and how to install it

A content lead has pasted the same brand guide into Claude forty times this quarter. Sentence case in headlines. "Customers", never "users". Every statistic linked to its source. Some weeks she forgets a paragraph, and the drafts drift.

A skill is where that guide should live. Anthropic defines skills as "folders of instructions, scripts, and resources" that Claude loads only when relevant (Anthropic Help Centre, What are skills). Claude sees each skill's name and description up front and reads the rest only when your request matches, so a team can install several without cluttering every chat.

The practical facts: skills work on Free, Pro, Max, Team and Enterprise and need code execution switched on; Team and Enterprise Owners can provision a skill to everyone (Anthropic Help Centre). Agent Skills is an open standard, and ChatGPT reads the same SKILL.md format (ChatGPT Learn, Skills).

Brand-voice skills are the obvious first build. We'd start with the monthly report instead, because a report that changes its definitions every month does more damage than a slightly off-brand sentence. Here is a complete skill. Northwind is made up; the rules are the ones we'd use.

The folder:

paid-media-report/
├── SKILL.md
└── templates/
    └── report.md

SKILL.md:

---
name: paid-media-report
description: Builds Northwind's monthly paid media report from Google Ads, GA4 and HubSpot. Use for ad performance reports, ROAS questions or channel comparisons.
---

# Paid media report

## Before calculating anything
1. List the data you have: source (Google Ads, GA4, HubSpot or an
   uploaded export), date range, currency and row count.
2. If a source is missing, say so and leave that section out.
   Never fill a gap with an estimate.

## Definitions (don't change these unless asked)
- ROAS = total conversion value / total spend for the period, shown as
  a ratio to one decimal place (4.2x). Compute from totals. Never
  average daily ROAS.
- Label every platform figure with its attribution setting: the Google
  Ads conversion window, the GA4 lookback. If you can't read the
  setting, write [WINDOW UNKNOWN].
- CRM revenue = HubSpot closed-won amount for deals whose original
  source matches the campaign, in GBP. Don't convert currencies; list
  other-currency deals separately.
- HubSpot searches return at most 200 results a page. Fetch every page
  before totalling.
- Bounce rate is GA4's (sessions not engaged). Never compare it with
  Universal Analytics figures.

## Calculating
- Use code for every calculation and include it.
- Show platform ROAS and CRM ROAS side by side for each channel. The
  gap is a finding, not an error to smooth over.

## Writing the report
- Follow templates/report.md.
- Tag every sentence [DATA], [CALC] or [ASSUMPTION]. An explanation for
  a change is [ASSUMPTION] unless the data shows the cause.

## Before finishing
- Recheck total spend against the figure you read at the start.
- List campaigns with no matched deals, and deals with no campaign.
- End with a section called "What this rests on" listing every
  assumption.

Three choices worth copying. The description says when to use the skill in the words your team says ("ROAS questions"), because that's what Claude matches. The definitions come from the platforms' own docs: Google Ads defines ROAS as "total conversion value divided by total spend" (Google Ads Help glossary), and HubSpot documents the 200-result page limit and the lack of currency conversion (HubSpot Developers). And the last section forces the model to list its assumptions, which is where a report usually goes wrong.

1. Switch on code execution. Skills need it (Anthropic Help Centre, What are skills).
2. Build the folder. The folder name matches the skill's name. The name is up to 64 characters; Claude's help centre caps the description at 200 characters for uploaded skills (How to create custom skills).
3. Zip the folder, not its contents. The zip should contain paid-media-report/ at its root. The help centre marks files loose in the zip root as the wrong layout.
4. Upload it. Go to Customize > Skills, add the zip, and switch it on. On Team or Enterprise, an Owner can provision it so it appears for every member.
5. Connect the data. The skill assumes Claude can read Google Ads, GA4 and HubSpot. Setup for each, with official links, is in AI for marketing teams.
6. Test it. Try several prompts that should trigger it ("how did paid search do last month?", "ROAS by channel for Q3") and check Claude's thinking shows it loading the skill. If it doesn't fire, rewrite the description in the words that failed.

In Claude Code, Anthropic's marketing plugin installs with two commands:

claude plugin marketplace add anthropics/knowledge-work-plugins
claude plugin install marketing@knowledge-work-plugins

Anthropic does publish marketing skills, though not where most people look.

Its knowledge-work-plugins repository includes a marketing plugin, "primarily designed for Cowork" and also usable in Claude Code. It bundles five skills (content creation, campaign planning, brand voice, competitive analysis, performance analytics), seven commands including /draft-content, /brand-review, /performance-report and /seo-audit, and suggested connectors: Slack, Canva, Figma, HubSpot, Amplitude, Notion, Ahrefs, Similarweb, Klaviyo and Supermetrics (Anthropic, marketing plugin README). The repository calls its plugins "generic starting points".

That caveat matters. If no brand style guide is configured, /brand-review asks about your guidelines or gives "a generic review" for clarity, consistency and professionalism. Without your guide, it checks against an average. The same goes for performance-analytics: it can't know your attribution window.

Our view: install the plugin to see how it's built, then replace its brand-voice and performance skills with your own within a week. A brand-voice skill follows the same pattern as the report above: rules Claude can check ("no exclamation marks"), not adjectives ("friendly but professional").

A skill makes Claude consistent. Consistent and correct are different things.

• The skill doesn't load. Ask for "a quick post for the socials" and a voice skill described in terms of "LinkedIn and X" may not fire, so Claude improvises the voice it was meant to fix.
• It repeats a mistake faithfully. If the report skill has the wrong attribution window, every report is wrong, for everyone it's provisioned to.
• It can't check facts. A voice skill makes an invented statistic sound exactly like you. A 2023 study found 18% of GPT-4's citations were fabricated (Walters and Wilder, Scientific Reports); fluent and sourced are separate properties.

Skills can also run code, and Anthropic's advice is to "stick to trusted sources" (Introducing Agent Skills). Read a third-party skill's instructions and scripts before installing it.

A two-minute check for a marketing skill

1. Ask which skill it used. If yours isn't named, sharpen its description.
2. Run a known month. Compare the skill's report with one you've already checked.
3. Recompute one figure from the totals it shows.
4. Ask: "Which statements in this report aren't supported by the data you read?"

Steps 2 and 3 test the skill. Step 4 tests a particular answer, and it's what Count does: connect it to Claude, ask it to audit the report, and each claim comes back labelled a citation, an assumption, model knowledge or a hallucination, with its reasoning chain.

Are there official Claude skills for marketing?

How do I use Claude for marketing?

Can I use Claude skills in ChatGPT?