Claude skills for marketing: the 6 we run our outreach on (free to copy)

Anthropic's help page calls skills "folders of instructions, scripts, and resources that Claude loads dynamically". In practice a skill is a folder with one SKILL.md file. The top of the file holds a name and a one-line description of when to use it. Claude reads only those descriptions until a task matches, then loads the full file and any reference files beside it.

Why that matters for marketing:

• Your rules apply every time. No more pasting the same brief into every chat.
• They cost little when idle. Anthropic's Claude Code docs suggest moving task-specific instructions out of CLAUDE.md and into skills, because skills load only when used.
• They work in more than one place. Skills run in Claude Code, in the Claude apps on Free, Pro, Max, Ultra, Team and Enterprise with code execution on, and in beta on the API.

A plugin is a bundle: skills, connectors and sub-agents installed together. Anthropic's own marketing plugin has seven commands (draft content, campaign plan, brand review, competitive brief, performance report, SEO audit, email sequence) and is built mainly for Cowork.

This is the only skill we wrote. It is short on purpose. The skill file tells Claude what to read before doing any marketing work, which craft skill handles which job, and the rules that can't bend. The detail lives in one knowledge-base file it points to:

• Who we sell for: the offer, the countries, the entry price.
• Approved claims: the only proof lines a draft may use without asking. Any other number or client name needs Mark's OK.
• Principles, each with its outside source: start from clients who already paid; the free working sample is the outreach; one variable per test.
• A weekly loop: what runs daily, Monday, Wednesday and Friday.
• Anti-patterns: every correction Mark has made, dated, so the same mistake doesn't come back.

To copy the pattern, make a folder called marketing-director inside .claude/skills, with a SKILL.md that has four parts: a description listing every kind of marketing task, a "load first" list of your own files, a table routing each job to a craft skill, and three to five hard rules. Ours include "nothing reaches a prospect without approval" and "no verdict below 30 delivered".

Corey Haines' cold-email skill writes and reviews outbound: first emails, follow-ups, plays, reply handling and sending setup. It ships with ten reference files, from subject lines to deliverability. The rules we lean on most, from its SKILL.md:

• Write like a peer, not a sales machine.
• Personalization must connect to the problem. If you can delete the opener and the email still makes sense, it isn't working.
• Ask for interest, not a 30-minute call.
• Subject lines of two to four words, lowercase.
• Send from secondary domains, verify the list first, and turn off open tracking.
• No AI tells: no "it's not X, it's Y", no self-answered questions, no em dashes.
• One reply stops every channel.

We add our own limits on top: plain text, under 90 words, no link in the first email, signed with a first name.

Also from Haines' library. It runs research in three modes: what you already have (transcripts, surveys, reviews, support tickets), what's online (Reddit, review sites, forums, communities), and asking directly (interviews and surveys). It turns the findings into themes ranked by how often and how strongly people say them.

Two rules make it worth installing. It won't draw messaging conclusions from fewer than five independent data points per segment. And when a persona detail is missing, it leaves the field blank instead of inventing one. It also treats fetched reviews and posts as untrusted data, so text inside a page can't steer it.

We run it as step 0 before any new niche gets a list or a template.

From Growth Engine X's coldoutboundskills. Give it a website or a business brief and it produces at least 15–20 campaign ideas, from broad lists to narrow ones. Every idea has two levers: who is on the list, and what value it leads with. The narrower the list, the more the message can name the reason they were picked.

Three campaign types are required in every plan: a creative-ideas campaign, a new-hire campaign for people new in their role, and a lookalike of your best customers. Its test for each idea: "If a sales rep had 10 minutes to research a company before reaching out, what would they look for and why?"

The "I had a few ideas for your business" email is easy to ruin. Ask a model for three ideas about a company and it invents. This playbook fixes that with slot discipline: you name what bullets 1, 2 and 3 are about, once, for the whole campaign. The model only fills each slot with a fact true of that company.

The repo reports its own test: free-form ideas were usable 2 times out of 5; slot-defined ideas 21 of 23 (91%). One hard gate comes with it: a person writes three complete examples before any model call, and the model is never graded against examples it drafted itself.

In the current repo the playbook is named creative-ideas under skills/playbooks/. We copied it when it was called playbook-creative-ideas, and kept that name.

The last Growth Engine X skill stops the most common mistake in outbound: changing the list, the copy and the offer at once, then not knowing which one worked.

Each test changes one thing: the list, or the copy. It names a hypothesis, the single variable and the pass/fail line before launch, launches both arms the same day, and reads results at day 21. It also sizes the test. At a 1% positive reply rate, spotting a 2× lift needs about 500 sends per arm; a 1.5× lift needs about 2,000. Before any test, your baseline must reach 1% replies after 200 sends, or the problem is your setup, not your copy.

We run smaller than that, so we add a floor: no verdict below 30 delivered, and every result is reported with its sample size and date.

• A skill can't fix the offer or the list. The cold-email skill makes a good email sharper. It can't make the wrong buyer want the wrong thing. The practitioners behind our brain agree: the list and the offer move results more than the wording.
• Outside skills assume outside tools. The cold-outbound skills are written around Smartlead, Clay and paid data. We map each step to our own sender and scrapers. Read the SKILL.md before you run it, and change the tool names to yours.
• Big packs aren't better. In our review, libraries with hundreds of skills were broad and shallow. Six that match your actual jobs beat 400 you never open.
• Skills don't send anything. They decide and draft. Sending needs a mailbox, a sender and a person approving. Keep it that way for anything that reaches a prospect.
• Your voice is still yours. A skill can enforce rules. The offer, the template and the first three examples should come from you; the creative-ideas playbook insists on it.
• Claude doesn't make images. For ads and visuals you need an image or video model. Our ad-creative work at Tala used Claude agents to build pipelines that call video models.

1. 1.In your project, make a .claude/skills/ folder.
2. 2.Install Haines' library with npx skills add coreyhaines31/marketingskills, or copy just the cold-email and customer-research folders from GitHub.
3. 3.Clone Growth Engine X's coldoutboundskills and copy campaign-strategy, experiment-design and the creative-ideas playbook. Keep each LICENSE file.
4. 4.Write your own marketing-director/SKILL.md: the description, a "load first" list, the routing table, your hard rules. Put the long material (claims, ideal customer, anti-patterns) in a separate file it points to.
5. 5.Ask Claude Code for a real task, such as "write the first email for this segment", and check that it loads the brain first.
6. 6.Every time you correct its output, add one dated line to the anti-patterns. That file is where the value builds up.

New to the tools? Claude chat vs Cowork vs Claude Code explains where skills run, and Claude for small business covers the rest of what Claude can do for you.