Marketing skills for AI agents: teach Claude to do growth work Write the job down once, then let the agent repeat it.

marketing skills for AI agents are instruction files that teach an agent like Claude to do one growth job your way, every time. each skill is a folder with a SKILL.md file: a name, a description of when to use it, then the steps, output format and rules. write the job down once and the agent repeats it.

What a skill is

a skill is a folder with a SKILL.md file inside. the file starts with a short header holding a name and a description, then plain instructions. Claude reads only the name and description until a task matches, then loads the full file. it turns a job you explain every week into a job you explain once.

the details, from Anthropic’s Agent Skills overview:

• name: up to 64 characters, lowercase letters, numbers and hyphens only
• description: up to 1,024 characters, and it must say both what the skill does and when to use it. this is what Claude matches your request against, so it matters most
• body: the steps, examples and rules. it only enters the conversation when the skill is used
• extra files: templates, reference notes or scripts can sit in the same folder and load only when needed

where skills run: in Claude Code, skills live in ~/.claude/skills/ for you or .claude/skills/ for a project, and you can call one directly with /skill-name (Claude Code skills docs). in claude.ai you upload a zip in settings, on paid plans with code execution turned on. the API has its own upload. skills do not sync between these, so add them to each place you use. Claude Code skills also follow the open Agent Skills standard, which other AI tools support.

Marketing skills repos on GitHub

the repository people usually mean when they search “marketing skills for AI agents Corey Haines” is coreyhaines31/marketingskills. it is an MIT-licensed collection of marketing skills covering copywriting, CRO, ads, ad creative, ASO, analytics and more, installable in Claude Code and other agents that support the Agent Skills format.

the other useful place to look is Anthropic’s own skills repository, which holds example skills, the spec and a template to copy.

use public skills as starting points, not finished tools. a generic ad creative skill does not know your kill rule, your target cost per trial or the phrases your users say. read it, keep the structure you like, and rewrite the rules with your own numbers.

also read every file before you install anything. Anthropic’s docs warn that a skill can direct Claude to run tools or code, so only use skills from sources you trust, and audit the rest like you would any software.

Which growth jobs make good skills

good skills are jobs that repeat, have clear inputs, and end in a fixed output you can check. weekly reviews, report summaries and first drafts from real user words all fit. bad skills are jobs that need taste in the moment, touch money without a human, or speak for your brand in public.

good skill	why it works	bad skill	why it fails
weekly creative test review	same inputs, written kill rule, fixed table out	replying to Reddit or TikTok comments	needs a person who read the thread
mining reviews for hooks	clear input, quotes are checkable	pausing or scaling ads on its own	money moves without a human
weekly metrics summary	same export every week	brand strategy for next year	one-off judgment, no repeatable steps
first draft of a creative brief	template plus last week’s winners	anything with no written rules yet	you cannot teach what you have not written

a simple test: if you could hand the job to a new hire with a one-page doc and check their work in five minutes, it is a good skill.

How to write one: weekly creative test review

start with a job you already do by hand. the weekly creative test review is a good first skill: you export results, apply your kill rule, and decide what to kill, hold, scale and spin into variants. the skill below does the sorting and drafts the decisions, and a person approves them.

when i was testing 150+ creatives a month, this review was the job that ate the most hours and needed the most discipline, because it is easy to keep a loser alive when you like it. a skill applies the written rule every time. the full method behind the rule is in app creative testing.

save this as .claude/skills/reviewing-creative-tests/SKILL.md and replace the values in curly brackets:

---
name: reviewing-creative-tests
description: Reviews a weekly export of app ad creative results, applies the team's written kill rule, and drafts kill, hold, scale and spin decisions for a human to approve. Use when the user shares a Meta, TikTok or Apple Ads creative export or asks for the weekly creative review.
---

# Weekly creative test review

## Inputs
- A CSV export with one row per creative: creative id, concept, hook, spend, installs, trials, day 7 retained users
- If any column is missing, stop and list what is missing. Do not estimate it.

## Rules
- Target cost per trial: {your target, e.g. $X}
- Minimum spend before judging a creative: {e.g. $Y}
- Judge on cost per trial and cost per day 7 retained user, never on cost per install alone.

## Steps
1. Check the file has every input column. Count the rows.
2. For each creative, calculate cost per install, cost per trial and cost per day 7 retained user. Show the math.
3. Mark any creative under the minimum spend as "not enough data". Do not judge it.
4. Apply the rules:
   - Kill: more than 20 percent above target cost per trial after minimum spend
   - Hold: within 20 percent above target, or spend just past the minimum
   - Scale: below target cost per trial and below target cost per retained user
5. For every Scale creative, suggest three variants: a new hook on the same body, the same script with a new creator, and the same idea as a screen recording.
6. Group results by concept and say which concepts keep winning.

## Output format
1. A table: creative, concept, hook, spend, cost per trial, cost per day 7 user, decision, one-line reason
2. "Concepts winning this week" with two to three bullets
3. "Variants to brief" as a numbered list
4. "Questions for the team" for anything unclear

## Guardrails
- Never pause, launch or change budgets. Draft decisions only. A person approves.
- Never invent numbers. If a value is missing, write "missing".
- Write in plain words. No hype, no exclamation marks.
- If results look wrong (for example trials higher than installs), flag it before deciding anything.

the “hold within 20 percent” line is an illustrative rule, not a benchmark. use the rule your team already trusts.

How to test a skill

test a skill by running it on past weeks where you already know the right answers. if its decisions match yours, it works. Anthropic’s skill authoring best practices say to build evaluations first: run the task without the skill, write three test cases, then add only the instructions that fix the gaps.

a practical test plan, about two hours:

1. baseline: give Claude last month’s export with no skill and your plain request. note what it gets wrong
2. three test weeks: pick three past weeks where you know what you killed and scaled, including one messy week with missing data
3. run the skill on all three. compare its decisions with yours, row by row
4. fix one thing at a time: if it judged on installs, make that rule more prominent. if it guessed a missing value, strengthen the guardrail
5. trigger test: ask for “the weekly creative review” in new words and check the skill loads. if not, sharpen the description
6. model test: if you use more than one Claude model, test on each, as Anthropic’s guide suggests

keep the three test weeks in the skill folder. every time you edit the skill, run them again.

How to keep brand voice human

skills should produce drafts and decisions for a person to approve, not public words that go out on their own. keep anything that speaks for your brand, like comments, community posts and replies, with a human who read the context. use skills to gather the raw material and to check the rules.

three habits that keep voice human:

• feed it real words: keep a file of exact phrases from reviews, support tickets and user calls, and tell the skill to quote from it, not invent
• ban your AI tells: add a short list of words and patterns your brand never uses, and ask the skill to flag them in drafts
• a human edits the last mile: every public line gets rewritten by a person before it ships

this is the line between useful automation and spam, and it is the core idea in agentic growth marketing. for how skills fit into a bigger agent that runs the weekly growth loop, see the AI growth marketing agent.

A two-week plan to ship your first skill

you can ship a first marketing skill in two weeks with a few hours of work. week one is writing down how you already do the job. week two is testing and tightening. the goal is one skill you trust enough to run every monday, not a folder of twenty you never use.

day	do this	time
week 1, day 1	do the creative review by hand, writing every rule you apply	1 hour
week 1, day 2	paste the SKILL.md above and fill in your own rules	30 min
week 1, day 3	pull three past weeks as test cases	30 min
week 1, day 4 to 5	run the baseline and the skill, compare decisions	1 hour
week 2, day 1	fix the biggest gap, rerun all three tests	45 min
week 2, day 2	run it on this week’s real data, a person approves every decision	30 min
week 2, day 5	write down what you still had to fix by hand. that is your next edit	15 min