Claude Code for Marketers: Set Up Agents, Skills and Automations Without Being a Developer

Why a “coding” tool is the best marketing agent

Chat tools are stateless. Every session, you re-paste your brand guidelines, re-explain your audience, re-describe the report format, and hope. Claude Code inverts that: it operates inside a folder on your machine, and everything in that folder is context it can read, write and remember.

For a marketer, that folder is not a codebase. It looks like this:

marketing/
├── CLAUDE.md                 # who we are, how we write, what we sell
├── brand/
│   ├── tone-of-voice.md
│   ├── offers-and-pricing.md
│   └── never-say.md
├── data/
│   ├── ga4-exports/
│   └── search-console/
├── campaigns/
│   └── 2026-q3-launch/
├── reports/                  # the agent writes here
└── .claude/
    ├── skills/               # repeatable procedures
    ├── agents/               # specialist subagents
    └── settings.json         # permissions and guardrails

Three properties make this work far better than a chat window: the agent can read your actual files, it can write finished deliverables instead of chat messages, and the instructions are version-controlled and reusable instead of living in someone’s scroll history.

You are not teaching an AI to do marketing. You are writing down how your marketing works — which you needed to do anyway.

Setup in 15 minutes

Claude Code runs in the terminal, and also ships as a desktop app, a web app and extensions for VS Code and JetBrains. If the terminal makes you nervous, start with the desktop or IDE version — everything below works identically.

1. Install it — the current install commands and system requirements are on the Claude Code overview page. The npm route is npm install -g @anthropic-ai/claude-code.
2. Create your marketing folder and put one real thing in it — last month’s report, your brand guidelines PDF, anything.
3. Open a terminal in that folder and run claude.
4. Type: “Read everything in this folder and tell me what you understand about this business.”

That last step is the real test. Whatever it gets wrong is exactly what is missing from your documentation — which is what you fix in the next step.

Step 1 — CLAUDE.md, your brand brain

CLAUDE.md sits at the root of your folder and is loaded automatically at the start of every session. It is the difference between generic AI output and output that sounds like you. Spend real time here; everything else compounds off it.

A working starting template:

# About this business

We are a [B2B / D2C / local service] business selling [what] to [who].
Primary markets: India (Bangalore, Mumbai, Delhi NCR) and the US (East Coast).
Average deal size: [x]. Sales cycle: [x weeks].

# Who we are for

Ideal customer: [role, company size, trigger event that makes them buy]
Not our customer: [who we politely decline — be specific]

# What we sell

1. [Offer one] — [price band] — [the outcome the client buys]
2. [Offer two] — [price band] — [the outcome the client buys]

# How we write

- Second person, direct, short sentences.
- Concrete numbers over adjectives.
- We explain the mechanism, we never say "cutting-edge" or "revolutionary".
- Indian English spelling, INR as Rs. for India, USD for US pages.

# Words we never use

hustle, synergy, unlock, game-changer, leverage (as a verb), "in today's
fast-paced world", any claim we cannot evidence.

# Non-negotiable rules

- Never invent statistics, case study numbers or client names.
- Never state a price that is not in offers-and-pricing.md.
- If a claim needs a source and you do not have one, flag it instead of writing it.
- Anything that spends money, goes public, or makes a promise needs my approval.

# Key links

Site: https://example.com
Analytics property: [GA4 property id]
Main conversion: WhatsApp click + enquiry form

Step 2 — Your first five marketing skills

A skill is a folder containing a SKILL.md file that describes a repeatable procedure. It has a name and a description in its frontmatter, and Claude loads it automatically when your request matches that description. You write the procedure once; you never explain it again.

The structure:

.claude/skills/weekly-report/SKILL.md

And the file itself:

---
name: weekly-report
description: Produce the Monday marketing performance report. Use whenever
  asked for the weekly report, a performance summary, or "how did last
  week go".
---

# Weekly Marketing Report

## Inputs
- GA4: sessions, engaged sessions, conversions, by channel
- Search Console: clicks, impressions, top 20 queries, top 10 pages
- Any campaign notes in campaigns/ for the period

## Steps
1. Pull this week (Mon-Sun) and the previous week.
2. Calculate week-over-week change for every metric.
3. Flag ONLY changes greater than 15% or 20 conversions. Ignore noise.
4. For each flagged change, state the most likely cause and say how
   confident you are. Never assert causation you cannot support.
5. Write the three actions worth taking this week, ranked by impact.

## Output
Write to reports/YYYY-MM-DD-weekly.md using this structure:
- Headline: one sentence on the week
- The numbers: a small table
- What moved and why
- Three actions, ranked
- What I could not verify

## Rules
- Never fabricate a number. If a source is unavailable, say so explicitly.
- Round percentages to whole numbers.
- Maximum one page. This gets read on a phone.

The five worth building first, in this order:

1. weekly-report — the one above. Highest time saving, easiest to verify.
2. landing-page-qa — checks a URL against your pre-launch checklist and returns pass or fail per item.
3. seo-sweep — missing titles and metadata, broken internal links, images without alt text, orphan pages.
4. repurpose — one asset into LinkedIn, email and short-form scripts, all against tone-of-voice.md.
5. lead-reply — drafts a first response to an enquiry in your voice. Drafts only. Never sends.

Write them one at a time, and use each one for a week before writing the next. A skill you have not stress-tested is a liability, not an asset.

Step 3 — Subagents for the heavy jobs

A subagent is a separate worker with its own context window and its own tool permissions. It goes away, does a bulky job, and reports back a summary — without flooding your main conversation with a thousand lines of raw data.

Defined as a file in .claude/agents/:

---
name: seo-auditor
description: Crawls a site and reports technical SEO issues. Use for site
  audits, pre-launch checks, or "why isn't this page ranking".
tools: Read, Grep, Glob, WebFetch
---

You are a technical SEO auditor. You report findings; you never edit files.

For each page you check:
- Title present, unique, under 60 characters
- Meta description present, unique, 120-158 characters
- Exactly one H1, and it matches search intent
- Canonical present and self-referencing unless intentionally not
- All images have meaningful alt text
- No broken internal links
- Page is reachable from the sitemap and from at least one other page

Return a table sorted by severity. For every issue give the fix as a
concrete instruction, not a principle. If you cannot verify something,
say "unable to verify" — never guess.

Note tools: Read, Grep, Glob, WebFetch. This subagent physically cannot edit your files — it has no write tool. That is a guardrail enforced by the harness, not by politely asking the model. Use this pattern for anything that touches production.

Skill or subagent?

• Skill when you want a procedure followed in your current conversation.
• Subagent when you want bulk work done elsewhere and only the conclusion brought back — or when you need to hard-restrict what it can touch.

Step 4 — Connect real data with MCP

Everything so far runs on files. The step change comes when the agent can read your live data. That is what MCP — the Model Context Protocol — is for. MCP servers expose external tools and data to the agent through a standard interface: analytics, ad platforms, CMSs, spreadsheets, CRMs, databases.

You add one once:

claude mcp add my-analytics -- npx some-analytics-mcp-server

After that, “pull last week’s organic traffic by landing page” is a sentence, not a CSV export. The connectors most marketing teams want first:

• Google Analytics 4 — sessions, conversions, channel and landing-page breakdowns.
• Google Search Console — queries, impressions, click-through rate, indexing status.
• Your ad platform — Meta and Google Ads, for spend, search terms and creative performance.
• Your CMS — so audits can read live pages and drafts can be created for review.
• PageSpeed / Core Web Vitals — so page performance checks are real measurements, not opinions.

Step 5 — Put it on a schedule

An agent you have to remember to run is a tool. An agent that runs itself is infrastructure. Headless mode lets you run a prompt non-interactively:

claude -p "Run the weekly-report skill for last week and save the output to reports/"

Wrap that in a scheduled task and the report is on your desk before you open your laptop:

• Mac or Linux: a cron entry, for example every Monday at 08:00.
• Windows: Task Scheduler, running the same command.
• Server or CI: a scheduled job in whatever runner you already use.

Good candidates for a schedule: the Monday report, a Friday SEO sweep, a daily broken-link check, a weekly competitor snapshot. Bad candidates: anything that publishes, replies or spends.

Step 6 — Guardrails that keep you safe

Three layers, in order of how much they matter.

1. Permissions

Claude Code asks before taking consequential actions, and you can pre-approve or block specific tools and commands in .claude/settings.json. Allow the boring reads. Never blanket-allow everything because the prompts annoy you — that prompt is the last thing standing between a typo and your live site.

2. Hooks

Hooks run your own commands automatically at defined points — before a tool runs, after it runs, when a session ends. Marketing uses: log every file the agent writes, block edits to your live templates, or run a link checker automatically after any content file changes.

3. The one-sentence rule

Put this in CLAUDE.md verbatim:

If an action spends money, goes public, or makes a promise — stop and ask me first.

It is not a technical control, and it is still the most valuable line in the file. Combine it with the tool restrictions above and you have both a policy and an enforcement mechanism.

Five mistakes to avoid

1. Skipping CLAUDE.md. Everyone does this and then complains the output is generic. The output is generic because the input was.
2. Building ten skills before using one. Ship one, use it for a week, fix what annoyed you. Then build the next.
3. Automating a broken process. If your reporting is wrong manually, you have just built a machine that is wrong faster.
4. Granting write access on day one. Read-only for a month. Every regret story starts with skipping this.
5. Not measuring. Log hours before and after from day one, or in six months you will have a strong feeling and no case.

On that last point — the measurement setup is its own job, and it is covered in How to Track AI Marketing Automation So It Shows Up in Revenue. If you are still deciding between tools, start with Claude Code vs Codex vs chat assistants.

Frequently asked questions