Claude Code for Marketing: What to Build First and How to Run Paid Media From It

When an ad-variation request lacks buyer context, approved claims, and

brand examples, the drafts can be difficult to evaluate. Put those

inputs alongside past campaign data so Claude can use them while

drafting, and you can check the result against them.

Claude Code for marketing combines project files, reusable task

instructions, and connected tools. This guide covers setup, the

differences from Claude Chat and Cowork, four workflows to build, and

the controls needed before giving it access to paid-media accounts.

What Is Claude Code, and How Is It Different From Claude Chat and Cowork?

**Claude Code is Anthropic's agentic tool that works inside a folder:

it reads the files there, edits them, runs commands, and connects to

your other tools, from the terminal, an IDE, the Claude desktop app, or

a browser**. It needs a paid Claude plan or an Anthropic Console

account.

For marketing, the folder is the point: your brand guide, briefs, and

exports become working context instead of things you paste into a chat.

What it does Best marketing Plans

use

• -------------------------------------------------------------------------------------------------------- ----------------- ----------------- -------------------------------------------------------

**[Claude Chat]{.underline}** Conversation. In One-off drafting, Free and paid

the desktop brainstorming,

app's Chat tab, editing something

it has no access you paste in

to your files

[**[Claude An autonomous Multi-step, Pro, Max, Team, Enterprise

Cowork]{.underline}**](https://support.claude.com/en/articles/13345190-get-started-with-claude-cowork) agent in Claude file-based tasks

Desktop and on (decks,

the web that spreadsheets,

takes a task and research) with no

returns finished setup

work, with access

to local files on

desktop

Claude Code Works in a Repeatable [Pro, Max, Team,

project folder marketing Enterprise](https://claude.com/pricing),

you choose, with workflows, shared or Console API billing

CLAUDE.md skills, live data

instructions, through MCP

skills,

subagents, hooks,

MCP servers and

plugins

So which Claude is best for marketing? If you're weighing Claude for

marketing work in general, regular Claude Chat handles one-off writing,

and Cowork takes a file-based task with no setup. Claude Code fits when

you want the same task done the same way every week, with instructions

and skills you can share. Two caveats. Anthropic is folding Cowork into

the main Claude app on Pro and Max plans, so the "Chat" and "Cowork"

labels may not appear in your account. And Cowork loads skills and

connectors from Customize in the app rather than from Claude Code's

\~/.claude folder, so a skill you build in one doesn't automatically

appear in the other.

Why Marketers Are Using Claude Code

The shift is from "AI writes a draft" to "AI runs a workflow on my

files and tools." A chat conversation mostly works from what you paste

into it, while Claude Code starts from a folder that already holds your

positioning, past campaigns, and rules, and writes output in the format

the next tool expects.

In [Anthropic's January 2026 account of its own marketing

workflow](https://claude.com/resources/articles/how-anthropic-uses-claude-marketing),

Austin Lau built a custom /rsa command for responsive search ads without

prior coding experience. It uses skills for brand voice, product

accuracy, and Google Ads rules, then exports 15 headlines and 4

descriptions per ad to a CSV for manual review. Anthropic reports a

reduction from 30 minutes per ad to 30 seconds; that is its team's

reported result, not a forecast for your workflow.

The workflow separates four responsibilities:

• Context lives in files. Brand voice, product facts, and channel

rules were written once, with the product marketing and copywriting

teams.

• The repeatable instructions live in a command that the marketer can

invoke again.

• The output matches the destination. It's a CSV the ad platform

accepts, not prose to reformat.

• A person approves before anything ships.

Start by describing one repeatable task and its review criteria. Our

guide to [building your own marketing

agent](https://synterai.com/blog/how-to-build-ai-agents-for-marketing)

covers workflows beyond a folder of skills.

How to Set Up Claude Code for Marketing Work (No Coding Required)

Pick Your Surface: Desktop App, Terminal or IDE

1. Desktop app. Install Claude Desktop, sign in, open the Code tab,

choose Local, and select your project folder. Claude Code is

included in the app; a separate CLI installation is not required.

1. Terminal. To install Claude Code on macOS, Linux, or WSL, paste

the command below; on Windows PowerShell, the equivalent is irm

https://claude.ai/install.ps1 \| iex. Then move into your folder and

type claude, and it asks you to log in on first use.

1. IDE. If you already work in VS Code, the Claude Code extension

runs the same engine inside the editor.

curl -fsSL https://claude.ai/install.sh \| bash\

cd acme-marketing\

claude

The desktop app and terminal can share MCP servers and settings files.

Use the [installation

instructions](https://code.claude.com/docs/en/overview) if

you prefer the terminal or another supported surface.

Create One Project Folder per Brand or Client

Start with one project folder per brand or client to reduce accidental

context mixing. A folder organizes context but doesn't isolate it:

personal \~/.claude/CLAUDE.md instructions and CLAUDE.md files in parent

folders can load too; extra directories can be added, and user-scoped

MCP servers like Synter's work across projects. Review those shared

instructions and connected accounts when client separation matters.

acme-marketing/\

CLAUDE.md always-on instructions\

brand/voice-guide.md\

brand/approved-claims.md\

research/customer-interviews/\

data/ CSV exports: ads, CRM, analytics\

drafts/ where Claude writes output\

.claude/skills/ project skills, one folder each\

.claude/settings.json permission rules for this project

Skills you save under .claude/skills/ belong to that project, so a

teammate who copies or syncs the folder gets the same commands.

Write a CLAUDE.md With Brand Voice, ICP and Product Context

CLAUDE.md is a plain markdown file in the project root that Claude Code

reads at the start of every session. It is where your product marketing

context goes: who you sell to, what you sell, how you sound, and what

you never say. Running /init generates a starter file, though it is

designed around codebases, so for a marketing folder it is usually

faster to write it yourself or ask Claude to draft one from your brand

guide.

Keep CLAUDE.md concise, with concrete instructions you can check;

Anthropic recommends [fewer than 200 lines per

file](https://code.claude.com/docs/en/memory). A full voice

guide can live in a separate file referenced with an @ import.

\# Acme Analytics: marketing context\

\

\## Who we sell to (ICP)\

• RevOps and marketing ops leads at B2B SaaS companies, 50-500

employees\

• They own the reporting stack and get blamed when dashboards disagree\

\

\## Product\

• Acme reconciles ad platform, CRM and web analytics data into one

pipeline report\

• It does NOT do attribution modeling or run ads. Never imply it does.\

\

\## Brand voice\

• Plain, specific, a little dry. Short paragraphs. No exclamation

points.\

• Full guide: \@brand/voice-guide.md\

\

\## Rules\

• Only use claims listed in \@brand/approved-claims.md\
• Never invent statistics. If a number has no source file, leave a

\[NEEDS SOURCE\] marker\

• Never name competitors in ad copy\
• Write all output to drafts/ with the date in the file name

The "does NOT do" line identifies a product boundary to check in the

output. CLAUDE.md provides context, not enforcement, so review the draft

for violations of those instructions.

Your First Prompt (a Real One, With Expected Output)

Pick a task you know well enough to judge the answer:

Read CLAUDE.md and every file in research/customer-interviews/. List the

five pains buyers mention most, each with one verbatim quote and the

file it came from. Then write three LinkedIn ad hooks for the top pain,

under 150 characters each, in our voice. Save everything to drafts/.

Claude reads the interview files without asking, because reading inside

your working folder needs no approval, and in Manual mode it asks before

writing the file to drafts/. You should get a ranked list with quotes

and file names, then three hooks. Your job is the check: open two cited

files and confirm the quotes are real. If one is paraphrased, add a

CLAUDE.md rule ("quotes must be verbatim") and rerun. Correct once,

write it down: that loop is the whole method.

Skills, Subagents and MCP Servers: The Three Things That Make Claude Code Useful for Marketing

Choose the extension according to the work you need to repeat or

connect:

• You keep typing the same prompt to start a task: save it as a skill.

• A side task floods the conversation with output you won't reread:

route it through a subagent.

• You keep copying data from a browser tab Claude can't see: connect

that system as an MCP server.

Skills (Reusable Instructions for Tasks You Repeat)

A skill stores reusable instructions in SKILL.md. Claude can load it

when a request matches its description, or you can invoke it with

/skill-name. Custom slash commands have been merged into skills. In a

regular session, the description is available first and the [full

instructions load when

invoked](https://code.claude.com/docs/en/skills).

Start with a task you can review: a brand-voice check, a claims check

against approved-claims.md, or an email sequence that follows your

nurture rules. Anthropic's [Marketing

plugin](https://claude.com/plugins/marketing), built mainly

for Cowork and compatible with Claude Code, includes draft-content,

campaign-plan, brand-review, competitive-brief, performance-report and

seo-audit. Plugins bundle extensions such as skills, subagents, hooks

and MCP servers. Our [marketing-skills

roundup](https://synterai.com/blog/best-marketing-skills-for-ai-agents-2026)

lists options by platform and function.

Subagents (Parallel Specialists for Research, Copy and Analysis)

A subagent works in a [separate context

window](https://code.claude.com/docs/en/sub-agents) with

its own instructions, tool access and permissions, then returns results

to the main session. For competitor research, ask each worker to return

the same one-page template with source links, so the main session can

compare summaries without carrying every retrieved page.

MCP Servers (Live Connections to Your Tools and Data)

The [Model Context

Protocol](https://modelcontextprotocol.io/docs/getting-started/intro)

connects AI applications to external systems. In Claude Code, add a

remote server and complete browser sign-in through /mcp. The command

below [connects HubSpot at user

scope](https://code.claude.com/docs/en/mcp).

claude mcp add \--transport http hubspot \--scope user

https://mcp.hubspot.com/anthropic

Once connected, Claude reads the live system instead of last Tuesday's

CSV. Ad platforms work the same way: Synter, our own product, is one MCP

server covering [16 ad

platforms](https://synterai.com/integrations), and the paid

media section below shows how to set it up.

Marketing Workflows Worth Building in Claude Code

Each workflow below has four parts: input files, a skill, an output in a

usable format, and a human check before anything leaves the folder.

Workflow Input Skill you write Output Human check

(example name)

• -------------- -------------- ---------------- -------------- --------------

Content and Page list, /content-audit Ranked fix Every changed

SEO audit search list plus fact against

performance drafted its source

export, refreshes

CLAUDE.md

Competitive Competitor Subagent per One-page brief Spot-check

intel brief URLs, your competitor with sources claims; no

positioning and dates copy reuse

Ad copy and Current ads /ad-variants Upload-ready Policy,

A/B test plan export, CSV plus test claims, one

performance, plan variable per

approved test

claims

Weekly Ad exports or /weekly-report Report in a Date range,

performance MCP data, CRM fixed template timezone,

report export, conversion

analytics definition

Content and SEO Audits

Drop the search performance export for your blog posts and landing pages

into data/. A /content-audit skill tells Claude what to flag: posts

losing clicks, product claims that contradict CLAUDE.md, missing

internal links between related blog posts, and pages competing for the

same query. Ask for a ranked table first, then refreshed drafts for the

top three. Claude Code has built-in web search and fetch tools, so it

can check whether a cited statistic still says what your post claims.

The human check: open the source for every changed number, and review

internal linking suggestions against what your content strategy is

pushing this quarter.

Competitive Intel Briefs

Give Claude competitor URLs and your positioning file, then ask for one

subagent per competitor. Use a consistent output template: target

audience, published product claims, changes since the last brief, and a

source link for each point. Merge those briefs into a dated comparison

and check the claims before using them in copy.

Ad Copy Variations and A/B Test Plans

For an /ad-variants skill, supply the platform's character limits,

approved claims and current ads export with performance data. Ask for a

CSV in the platform's import layout and a test plan that changes one

variable at a time. Review policy compliance and claims before upload.

Our guide to [automating creative

tests](https://synterai.com/blog/creative-testing-automation-guide)

covers multi-armed bandit testing and fatigue detection for ongoing

testing.

Weekly Performance Reports From CRM and Campaign Data

Define the reporting window, timezone, conversion event and system of

record in a /weekly-report skill. Use CRM closed-won data for revenue

reporting, while retaining platform data for campaign analysis. Begin

with CSV exports, then connect live sources. [Local Desktop scheduled

tasks](https://code.claude.com/docs/en/desktop-scheduled-tasks)

can read those exports while the app is open and the computer is awake.

Cloud routines

run remotely from repositories and connectors; included connector tools

can run without approval prompts, so restrict reporting jobs to

read-only access.

Running Paid Media From Claude Code With an MCP Server

Exports support retrospective analysis, but budget-pacing decisions need

a current reporting window. A live connection lets Claude retrieve

account data and request changes. Configure permissions and spending

limits before enabling those writes.

Some ad platforms offer their own MCP servers, each scoped to one

platform and some read-only; our [comparison of official Google and

Meta servers with a cross-platform

one](https://synterai.com/blog/synter-mcp-vs-google-ads-mcp-vs-meta-ads-mcp)

covers the trade-offs. The steps below use one server across platforms,

Synter, our own product.

Connect Ad Accounts Through One MCP Server (OAuth, No Keys in Chat)

[Add Synter's hosted

server](https://synterai.com/installation-instructions)

from a terminal with Claude Code installed. The hosted connection uses

browser sign-in and does not require a Synter package or API key.

claude mcp add \--transport http \--scope user synter-ads

https://mcp.synterai.com

1. Start Claude Code, open /mcp, select synter-ads, and complete Synter

sign-in in the browser. Claude Code stores the token, so nothing

gets pasted into chat.

1. Connect your ad accounts inside Synter. This is a separate

authorization: MCP OAuth lets Claude Code talk to Synter, and

advertiser OAuth lets Synter talk to Google or Meta, so connecting

the agent doesn't connect any ad account by itself.

1. Because \--scope user saves the server to your user settings, it's

available in every project folder on your machine, and in the

desktop app's Code tab too.

This is the one step in this guide that needs a terminal. The same

install page lists a custom connector route for the Claude desktop app;

if you take it, the permission rules below need that connector's tool

names instead.

Synter's optional Claude Code plugin adds packaged paid media skills,

but the hosted connection is enough to start, and since the plugin may

add its own MCP configuration, check /mcp for duplicates before

installing it.

Start Read-Only: Cross-Platform Analysis Prompts

Begin with analysis before enabling account changes. Ask Claude to show

each tool call's cost and to leave campaigns unchanged. Reads can

consume credits: get_credit_balance returns the balance, and

forecast_tool_cost returns a tool's live price. Try the following

illustrative prompts:

• "Compare CPA and spend velocity across Meta, TikTok, and Reddit over

the last 30 days."

• "Audit Google search terms for the last 14 days and flag

non-converting queries with \>\$50 spend as negative keywords."

• "Write the flagged search terms to drafts/negatives.csv with spend,

clicks, and the reason for each. Don't add anything to the account."

The third prompt produces a file for human review. Use a read-only

search-terms tool and include account-specific context in CLAUDE.md,

such as brand terms you bid on and competitor names you want to retain.

Check the proposed exclusions before applying them.

Move to Guarded Writes: Approvals, Budget Caps and the Audit Log

Control writes on both sides of the connection: Synter enforces

workspace policy, while Claude Code controls whether it asks before

calling a tool.

Every Synter write passes a [server-side safety

gate](https://synterai.com/blog/how-synter-safety-gates-protect-ad-spend),

with the decision recorded in the workspace audit log. Workspace policy

blocks writes that exceed configured monetary limits or the daily budget

ceiling; approval cannot override a block. Under the [default Suggest

level](https://docs.synterai.com/mcp/approval-model),

pauses and budget, bid or keyword updates run immediately and are

audited. Enables, resumes, deletes, campaign creation, creative sends

and unrecognized writes are held. In production as documented in October

2026, hosted-MCP held writes require approval in the Synter app, and new

campaigns are created paused by default.

On Claude Code's side, your confirmation prompt is the only

per-call check on the writes Synter runs immediately, and it runs on

your machine, so Synter can't enforce it. That matters because auto

mode, where a classifier decides instead of asking you, is the built-in

starting permission mode in the terminal on recent versions. Add

explicit ask rules for Synter's write tools to the project's

.claude/settings.json: Claude Code never auto-approves a tool matched by

an ask rule, in any mode, and the desktop app reads the same file.

{\

"permissions": {\

"ask": \[\

"mcp\_\_synter-ads\_\_pause_campaign",\

"mcp\_\_synter-ads\_\_enable_campaign",\

"mcp\_\_synter-ads\_\_update_campaign_budget",\

"mcp\_\_synter-ads\_\_execute"\

\]\

}\

}

Request Synter default Claude Code with the

(Suggest) rules above

• ----------------------- ----------------------- -----------------------

Pull performance or Runs; may cost credits Follows your permission

search terms mode

Pause a campaign Runs immediately, Asks you first

audited

Change a daily budget Runs immediately, Asks you first

audited, within caps

Resume a paused Held for approval in Asks you first, then

campaign Synter waits for approval

Create a campaign or Held; new campaigns Prompts only if that

send creative start paused tool is on your ask

list

The four-tool list covers the most common spend-changing calls; other

write tools, such as campaign plan and creative upload tools, need their

own entries. To make every call prompt, use "mcp\_\_synter-ads", which

matches all of the server's tools, reads included. For an unattended

job, Synter's docs describe Off as read-only but also ambiguously say

routine updates run under Off and Suggest. Confirm Off blocks pauses and

budget changes in production before connecting a scheduled job to

advertiser write access. Cloud routines call tools on included

connectors without asking; local ask rules won't protect them.

Limits, Costs and Risks to Plan For

Usage limits. Pro and Max plans have a five-hour session limit and a

weekly limit, shared across Claude and Claude Code. Every subagent sends

its own requests against the same limits, so parallel runs use the

allowance faster. Run /usage to see where you stand.

Cost. Claude Code is included in paid Claude plans. One trap: if an

ANTHROPIC_API_KEY environment variable is set, Claude Code bills that

API key instead of your subscription. Ad tools add their own metering.

Synter's Solo plan, a

developer license with one workspace and one developer seat, is \$20 a

month with \$20 in usage credits you claim each billing period;

commercial brand workspaces start at Scale. A read is 2 credits

(\$0.05), a campaign launch is 1,000 credits (\$25), and Synter takes no

percentage of ad spend.

Reviewing AI-generated output. The model will write a plausible

number with no source if you let it. Keep a "no source, no number"

rule in CLAUDE.md, check claims against your brand guidelines and

approved-claims file, and remember that CLAUDE.md and skills are

requests. For checks that must run at a set point, use a tested command

hook, such as a script before a tool call configured to block it when

the check fails.

Permissions and change management. Before anyone connects a

production account, decide who owns the Synter workspace's autonomy

level, who reviews .claude/settings.json, and where approvals are

recorded. Keep the folder in version control or a shared drive so every

change to a skill or rule has an owner and a date.

FAQ

Which Claude is best for marketing?

Claude chat for one-off writing, Cowork for file-based tasks with no

setup, and Claude Code for workflows you want run the same way every

time, with CLAUDE.md context, skills, and live data through MCP. Cowork

needs a paid plan, Claude Code needs a paid plan or Console account, and

chat also works on the Free plan.

What are the Claude Code skills for marketing?

Skills are SKILL.md instruction files you run as slash commands or that

Claude loads when relevant. Useful marketing skills include brand-voice

and claims reviews, ad variant generators with platform character

limits, email sequence drafts, content audits, and weekly reports.

Anthropic's Marketing plugin adds ready-made ones such as campaign-plan

and brand-review.

How do you turn Claude into a marketing agency?

Claude Code can handle repeatable tasks such as research, drafts,

reports, exports and permitted account changes. People still choose

positioning and strategy, review client-facing work, and remain

accountable for spend. Assign an approval owner to each workflow.

Is there an AI for marketing that can actually run campaigns?

Yes, with limits. Through an MCP server, Claude Code can read live ad

accounts and make changes such as pausing campaigns or adjusting

budgets. Synter covers 16 ad platforms, with actions that vary by

platform API and permissions, and by default holds higher-risk writes

for human approval. Start read-only and add write access tool by tool.

Do you need to know how to code to use Claude Code for marketing?

No. Claude Code for marketers mostly means describing tasks in plain

English, and the desktop app needs no terminal at all. Learn a few

basics, such as folder paths, CSVs, and permission prompts, so you

understand what you're approving.

Conclusion

Use CLAUDE.md for brand context, skills for repeatable instructions, and

MCP for connected data. Before enabling ad-account writes, assign review

owners and configure the client's permission rules and Synter's

workspace controls.

Start with one project folder, a concise CLAUDE.md covering your ICP,

product limits and voice, and one weekly task saved as a skill. Review

its output before connecting live data. For paid media, [connect

Synter's MCP server](https://synterai.com/mcp) with the

command above and run a read-only cross-platform audit before enabling

writes. The MCP page currently offers a \$100 ad-credit voucher with

code MCP100.