# Claude Code for marketing: a project setup that works

How to set up Claude Code for marketing work: a project folder, a CLAUDE.md of brand rules, live data through one MCP connection, and three jobs to run first.

## Summary

Claude Code is built for developers, and marketers are using it anyway. The reason is simple. It reads and writes files on your machine, runs commands and connects to data tools, so a marketing job can end as a spreadsheet or a draft in a folder instead of a chat you scroll back through.

This post shows a setup for Claude Code for marketing. It covers the project folder, the one file that holds your rules, how to give Claude real data, three jobs worth running first, and the guardrails that keep a free-running agent from doing something you regret.

## Why use a coding agent for marketing work?

Because the output stays on your disk and the work repeats. A chat answer disappears into a thread. A Claude Code job writes a CSV, a brief or a page into a folder, in a format you control, and you can run the same job again next month with one prompt.

It also runs commands, so it can fetch data, transform it and save it in one go. For marketing that means the loop of pull, analyze, draft and file happens in one session. The trade-off is that it feels like a developer tool, because it is one.

## Claude Code for marketing: how do I set up a project?

Make one folder per brand or site, with a small fixed structure. A shape that works:

Research holds the data pulls, briefs hold one file per page you plan, and drafts hold the writing. Keeping the three apart stops a half-checked draft from sitting next to the facts it should cite.

Then add a `CLAUDE.md` file at the top. Claude Code reads it at the start of each session, so it is the right place for rules you do not want to repeat.

## What goes in the CLAUDE.md for a marketing project?

Put in what a new hire would need on day one, and keep it short. Five items cover most of it.

1. **Who we are and who we sell to.** Two sentences on the product and the reader.
2. **Voice rules.** Plain words, no hype words, sentence case headings, and your own banned list.
3. **Evidence rules.** Every number needs a source and a fetch time. No source, no number.
4. **Where files go.** Research to `research/`, briefs to `briefs/`, drafts to `drafts/`.
5. **Spend rules.** The most a single job may spend on data, and when to ask first.

The evidence rule matters most. Without it Claude fills gaps with plausible numbers. With it, the first thing a reader sees in a draft is a sourced fact or a plain "we do not know".

## Claude Code for marketing: how do I add live data?

Connect an MCP server that reaches the data, so the agent calls real endpoints instead of recalling numbers. One connection to looot covers search results, keyword volume, backlinks, company data and social data through one token.

The first use opens a browser sign-in. Top up before the first paid run, because a new workspace starts at zero and there is no trial credit. Searching and inspecting the catalog are free, so you can test the connection first. [The Claude MCP servers post](https://looot.ai/blog/claude-mcp-servers-for-data-work) explains how to vet a server and keep the list short.

## What does the data cost?

Each endpoint has a price per call, and you can read it before the agent runs it. The table shows the cheapest priced endpoints for two common marketing pulls, read from the live catalog when the page loads.

Marketing data price per call: live prices per call are read from the catalog when the page loads, see https://looot.ai/blog/claude-code-for-marketing.

- Keyword volume and CPC
- Google organic results

A keyword call returns up to a thousand keywords, so a research pass is a handful of calls. [What 1,000 lookups cost](https://looot.ai/blog/what-1000-lookups-cost) shows how to price a larger job.

## Which three jobs should I run first?

Start with jobs that produce a file you can check in five minutes.

**A keyword table.** Ask Claude to take thirty phrases around one topic, get volume and difficulty, and save `research/keywords.csv` with the source and time on every row. You can open the CSV and judge it at a glance.

**A competitor page comparison.** Give it your page and two competitor URLs. Ask for a table of the claims each page makes, the proof it shows and the call to action, saved to `research/page-compare.md`. The recipe for a [landing page teardown](https://looot.ai/recipes/landing-page-teardown) has the steps.

**A content brief.** From the keyword table, ask for a brief for one page: the question it answers, the sections, the facts to cite from the research folder and the internal links. Save it to `briefs/`. Writing the draft is the second step, and it should quote only from the research folder.

## What does a week of use look like?

Day one is setup and one small job: the keyword table. Day two you read the table, fix the prompt and rerun. By day three you have a brief and a first draft that quotes the research folder, and you edit it by hand.

The pattern repeats for each page. You spend the model's time on the pulling and sorting, and your own on choosing the topic and reading the draft. A page that took half a day of research now takes an hour of reading, and the research files stay in the folder for the next page.

Expect friction at first. The agent will misread an ambiguous brief, and the first CLAUDE.md will be too long or too vague. Treat each miss as a rule to add to the file, so the same mistake does not come back.

## What guardrails should I set?

Three, and set them before the first long job. Cap data spend in the prompt and in `CLAUDE.md`. Keep Claude Code's permission prompts on for commands that write outside the project folder, and read them instead of approving by habit. And never paste a token into a prompt or a file in the repo. Use the sign-in flow or an environment variable.

Review everything that leaves the folder. A draft that quotes the research folder is checkable. A draft that quotes nothing is a guess.

## What it cannot do

- It does not publish for you unless you connect and allow a tool that does. Keep publishing a human step.
- It does not run a job on a timer. A weekly refresh needs scheduled runs, not available yet, so rerun the prompt by hand.
- It does not know your market. It knows the data you pull and the notes you write.
- It does not make a weak offer strong. If the product story is thin, the draft will be thin.
- It will reread a long folder slowly. Keep research files small and specific.

Last checked 2026-10-03 against the looot skill file and the live catalog. Author: the looot team.

For the broader method, read [vibe marketing with an agent](https://looot.ai/blog/vibe-marketing-with-an-agent). For the SEO side, see [the AI SEO agent guide](https://looot.ai/blog/ai-seo-agent).

## Questions

### Claude Code for marketing: does it work?

Yes. It reads and writes files, runs commands and connects to data tools, so a marketing job can end as a CSV, a brief or a draft in a project folder. It is a developer tool, so expect a terminal.

### Claude Code for marketing: how do I set it up?

Make a project folder with research, briefs and drafts, add a CLAUDE.md with your voice, evidence and spend rules, and connect a data server such as looot. Then run one small job and check the file it writes.

### What marketing jobs work best in Claude Code?

Jobs with a checkable file as output: a keyword table, a competitor page comparison and a content brief. Run those first, then widen to more phrases or more competitors.

### How do I get real SEO data into Claude Code?

Connect an MCP server that reaches data endpoints. With looot, the agent searches the catalog for keyword volume or search results, reads the price, and calls the endpoint with one token.

### Claude Code for marketing: what does it cost?

You pay for the model, and you pay a price per data call. The live table on this page shows the cheapest keyword volume and search result endpoints. Put a spend cap in the prompt and in CLAUDE.md.

### Is it safe to give the agent access to my marketing accounts?

Give it the narrowest access that works. Start with read-only data tools, keep permission prompts on, and keep publishing a human step. Do not paste tokens into prompts or files.

## For agents

### Code (bash)

```bash
mkdir -p acme-marketing/{research,briefs,drafts}
cd acme-marketing
```

### Code (bash)

```bash
claude mcp add --transport http looot https://api.looot.ai/mcp
```

### First marketing job

```text
Use looot to build a keyword table for <topic> in this project.
1. Set up https://looot.ai/skill.md.
2. Search the catalog for keyword volume and show me the cheapest endpoint and its price before you call it.
3. Get volume and difficulty for 30 phrases around <topic>.
4. Save research/keywords.csv with the columns phrase, volume, difficulty, endpoint and fetched_at.
5. Add the 10 phrases with the best volume to difficulty balance to a short summary at the top of research/keywords.md, with one line on why each fits.
6. Report the total cost and stop if it goes over 50 cents.
```