Key takeaways
- Claude Code works where marketing data lives. It reads files, runs commands and reaches GA4, HubSpot and Google Ads through connectors.
- Build it in order: a rules file, read-only connectors, skills around your own tools, a knowledge base it reads first, then hooks.
- Reads run freely. Anything that spends money, changes an account or goes live waits for my go.
- The same setup ran a 12-agent pipeline that grew a target list from 441 to 2,795 accounts.
01What Claude Code does for a marketer
Claude Code is a coding agent that reads your files, runs commands and connects to tools like GA4, HubSpot and Google Ads, so a marketer can hand it whole jobs: pull the report, mine the search terms, audit the UTMs, research the accounts or build the landing page. You set the rules and approve anything that spends money, and it does the work in between.
This is the hands-on version. The wider system, with the knowledge base and 26 scheduled jobs, is in how I run growth marketing on Claude Code and Codex.
02Why a coding agent instead of a chat window
Most advice on Claude AI for marketing stops at prompts. I use Claude Code for growth marketing because the slow part of the job is the plumbing between the ad account, the CRM, analytics and the site. In a chat window, each of those trips is an export and a paste.
A coding agent makes those trips for you. Ask which search terms wasted money last month, and Claude Code pulls them from the API, drops the ones that converted, groups the waste by theme and stages negatives for approval in one request. It can rerun the query to check itself, split big jobs across subagents and read the same rules every session.
03The setup, in the order I'd build it
- A rules file.Claude Code reads CLAUDE.md at the start of every session. Mine says which tool to reach for, what needs my approval and how a session should close out. Codex reads AGENTS.md, so the rules live in one shared AGENTS.md that CLAUDE.md points to. Claude Code can also import it with an @AGENTS.md line.
- Connectors, read-only first.MCP, an open standard for connecting AI tools to other systems, is how Claude Code reaches your stack. Mine cover GA4, HubSpot, Semrush, LinkedIn Ads, Google Drive, Figma and Chrome. Add one at a time with
claude mcp addand start on read access. The docs warn that a server fetching outside content can expose you to prompt injection, so connect only servers you trust. - Skills around your own tools.Where a connector can't do the job, I wrote small command-line tools: gads for Google Ads, hs for HubSpot, utm for campaign tagging, enrich for work emails and recall for past sessions. Each one is wrapped in a skill, a folder with a SKILL.md file. Together they replaced about $24K a year in software.
- A knowledge base it reads first.Agents forget between sessions and will guess an account ID if you let them. Every session starts from canonical facts and the notes earlier sessions left, and ends by writing its own. The design, with 184 facts and 5,163 session notes, is in the longer note.
- Hooks last.Hooks are commands Claude Code runs at fixed points, such as session start or just before a tool runs. Mine load the right topic dossier and recent activity when a session opens. The docs say CLAUDE.md is treated as context, so anything that must never happen belongs in a PreToolUse hook, which can block a tool call outright.
04What a marketing skill for Claude Code looks like
Most lists of marketing skills for Claude Code are prompt packs. The useful ones wrap a system you own and carry your judgment. Here is a simplified version of my weekly search-term review:
---
name: search-term-negatives
description: Find Google Ads search terms that spend
without converting and stage negatives for approval.
Use for search terms, negatives or wasted spend.
---
1. Read the account facts first. Never guess an ID.
2. Pull 30 days of search terms with no conversions:
gads --json search-terms list --days 30 --no-conversions
3. Keep competitor and integration terms. Those
campaigns exist to buy them.
4. Group the rest by theme. Propose exact or phrase
negatives, with the cost of each term.
5. Write a preview and stop. Wait for my go.
6. After the go, add them and log the undo.
Rules for writing them:
- The description does the routing. Claude loads the full skill only when the job matches its description, so lead with the task and the words people use to ask for it.
- Put the stop in the file. Step 5 is why this skill can't spend money by accident. For a skill that writes, you can also set
disable-model-invocation: trueso it runs only when you type/search-term-negatives. - Write down the judgment calls. Line 3 is the one a new hire would get wrong.
0510 marketing workflows I run with Claude Code
Every row comes from work at Pipe17 or Commas. The last column is the approval rule.
| Workflow | What I ask | What it touches | Needs my go |
|---|---|---|---|
| Search terms | Which search terms spent last month without converting? Stage negatives. | gads, the Google Ads API | Yes, before negatives go live |
| Ad rewrites | Rewrite the weakest ads using only claims we can prove. | gads | Yes, before they publish |
| Attribution | Which closed deals touched paid search in the last 6 months, and at what cost? | hs, gads | No, read-only |
| UTM audit | Find this quarter's utm_campaign values that are really one campaign. | utm, HubSpot | Only to rename CRM values |
| Account research | Score these 400 accounts for fit and find the likely champion. | Subagents, the web, the ICP file | Only to import results |
| Enrichment | Find work emails for this list and prep a HubSpot import. | enrich, hs | Yes, before the import |
| Funnel pages | Build a signup page variant from this brief and QA it in Chrome. | Site files, Figma, Chrome | Yes, before it goes live |
| Tracking fix | Conversions went to zero on Tuesday. Find out why. | GA4, Google Ads goals, Chrome | Only to change a goal or tag |
| AEO panel | Rerun the prompt panel and list the next pages to ship. | Panel results, Semrush, GA4 | Only to publish pages |
| Scorecard | Write the Friday scorecard: spend against qualified outcomes. | GA4, HubSpot, gads, Google Drive | No; I read it before leadership does |
The AEO row follows my AEO playbook. The research row ran at Pipe17 as a 12-agent pipeline that also classified 13,000+ companies for tiering.
06Guardrails that make it safe to hand over
Setting up the automation was the easy part. The guardrails are what let me hand real accounts to an agent.
- Spend needs a go. Budgets, bids, campaign status and account settings change only after I approve a staged preview.
- Writes start as previews. The default outcome of a bad instruction is a preview nobody approves.
- Every write has a rollback.
- Secrets stay out. Credentials are redacted before they reach a model.
- Senders have brakes. Daily caps, bounce stops and a kill switch.
Mistakes to avoid
- Connecting every tool on day one.
- Keeping account IDs and budgets in the chat. They belong in a facts file.
- Vague skill descriptions, so the skill never loads or loads for the wrong job.
- Judging the setup on hours saved instead of the cost of a qualified outcome.
Which jobs I hand to agents and which I keep is its own note: AI agents for marketing.
07Claude Code vs Codex vs ChatGPT for marketing
I run Claude Code and Codex side by side. Codex is my primary for long-running and scheduled work: 95.9 billion tokens and one autonomous task that ran 15 hours and 37 minutes. Claude Code gets interactive deep work, browser tasks and research that fans out.
| Tool | What it is | Where it fits | Watch for |
|---|---|---|---|
| Claude Code | A coding agent for the terminal, IDE, desktop or web, with MCP, skills, subagents and hooks | Interactive work on live data, browser tasks, research fan-out | Long research fills the context; push bulk work to subagents |
| Codex | OpenAI's coding agent, with AGENTS.md, skills and subagents | Long-running and scheduled work | Needs the same rules, tools and memory first |
| Claude Cowork | Claude's agent mode for knowledge work, in folders and tools you choose, with plugins | Marketers who want agents without a terminal | Connect tools read-only first |
| ChatGPT or Claude chat | A conversation in a browser tab | Drafts and quick questions | Without connectors, work ends in copy and paste |
On Claude vs ChatGPT for marketing, either writes a decent first draft. What shows up in results is the setup around the model: access to your systems, rules it reads every session and a stop before anything spends money.
08How to use Claude for marketing if you don't code
You don't need to write code. Claude Code writes and runs it. You need to describe the job clearly and check the numbers. A first week:
- Day 1: write the rules file.A single page: your role, your accounts, what needs approval and how a session ends.
- Day 2: connect GA4, read-only.Ask for last month's conversions by channel and check every number against GA4 itself.
- Day 3: turn that report into a skill.Ask Claude Code to write the SKILL.md. Next week it's one request.
- Day 4: start a facts file.Account IDs, budgets, conversion definitions and naming rules, so nothing gets guessed.
- Day 5: make one write, with a preview.A UTM fix or a few negatives. Approve it, confirm it landed and know how to undo it.
Then add one connector or skill a week. If a terminal is a hard no, Claude Cowork for marketing is the gentler start, with an official marketing plugin for content, campaign plans, competitive briefs and SEO audits. Move to Claude Code when you want your own tools in the loop.
FAQQuestions people ask
- How do I use Claude for marketing?
- Give it access to the work instead of pasting into a chat. Install Claude Code, write a short rules file with your accounts and approval limits, connect one tool read-only and ask for a report you can check against the source. Add skills and write access only after the reads are right.
- Do I need to code to use Claude Code for marketing?
- No. You describe the job in plain English and Claude Code writes and runs the code. Your part is clear instructions and checking the output against the source before you trust it.
- What are the best Claude skills for marketing?
- The ones that wrap your own systems and rules. Mine wrap command-line tools for Google Ads, HubSpot, UTMs, enrichment and session search. Public skill packs are fine for drafting and audits, but a skill that knows your accounts and approval limits does more of the job.
- Claude vs ChatGPT for marketing: which is better?
- For drafting copy, either works. What changes results is the setup around the model: whether it can reach your ad account, CRM and analytics, follow your rules every session and stop for approval before it spends money.
- Is Claude Cowork good for marketing?
- It's a good place to start if you don't want a terminal. Cowork is Claude's agent mode for knowledge work: it works in folders and tools you choose and takes plugins, including an official marketing plugin for content, campaign plans and SEO audits. Once you want to run your own tools and scripts, try Claude Code.
