How to Build Your Business's Marketing Brain with Claude
- Katya Tarapovskaia
- 2 days ago
- 5 min read

The problem: Random Acts of AI
Most marketing teams have a strategy doc. They wrote it during a planning offsite. It lives in a Google Drive folder nobody has opened since Q1.
Then AI arrived, and the problem doubled. Now we don't just do random acts of marketing, we do them faster. We generate more content, launch more campaigns, and produce more deliverables. But the strategic filter is still missing.
The fix isn't another doc. It's a system that holds your strategy and routes it into every piece of work - every brief, every campaign, every prioritisation call.
That's what a Marketing Brain is. A persistent, context-aware system that knows your business, your ICPs, your voice, and your no-go zones, and applies them automatically.
Step one: Know which Claude tool does what
This is where most marketers get confused. Claude isn't one tool, it's three, and they have very different jobs.
Claude Chat — Your "Thinking Partner"
Use Chat for the messy front-of-funnel work: angles, hooks, hypotheses, decisions.
Best for:
Cracking a positioning angle — paste competitor pages, ask for five sharper angles you can own
Pressure-testing a campaign — have it role-play your ICP and rip the idea apart
Turning raw notes into a sharp POV — drop call notes, get a contrarian narrative for LinkedIn
Sanity-checking a brief in 60 seconds — "Where will this confuse a busy CMO?" Then fix before sending
Chat is for thinking out loud. It's fast, conversational, and disposable. Think of it as the whiteboard session you never have to schedule.
Claude Cowork — Your "Work Partner"
Cowork is for execution. You describe the outcome, walk away, come back to a finished deliverable.
Best for:
ABM 1-pagers at scale — feed 10 target accounts, get tailored briefs with hooks per persona
Reports your CMO actually reads — GA + HubSpot data in, narrative report with "so what" on top
Competitive teardowns in an afternoon — 5 competitor URLs in, positioning gaps and angle map out
Editorial calendar for the quarter — strategy + ICPs in, 12-week calendar mapped to funnel stage
Cowork is for doing. If the task has a clear output and you could brief a sharp junior on it, it belongs here.
Claude Code — Your "Build Partner"
Code is where leverage compounds. Persistent agents. Project-level skills. Auto-loaded context. This is the layer that turns Claude from an assistant into an operating system.
Best for:
A custom AI marketing team — Strategist, Researcher, Writer, Editor, all agents loaded per project
An always-on competitor watcher — daily diff of 10 sites, Slack pings when positioning shifts
A lead enrichment + scoring engine — CSV in, enriched, scored, ICP-tagged accounts out
Brand voice as a reusable skill — encoded once, and every brief, post, and email runs through it
Code is for compounding. You invest setup time once, and every future task runs through a system that already knows who you are.
The gap most marketers miss: Cowork and Code don't share memory
Here's the detail that trips people up. Even when Cowork and Code are pointed at the same project folder, they don't automatically share context.
Capability | Cowork | Code |
Skills | Loaded manually per chat | Auto-loads project skills |
Agents | No saved AI team by default | Custom sub-agents, per project |
Web / APIs | Sandboxed — needs connectors | Talks to web + APIs natively |
Token burn | Heavier — visual UI layer | Leaner — text-only context |
Cowork won't see agents you built in Code unless you bridge them via a CLAUDE.md file. This is the most important architectural detail in the whole system, and it's the one nobody talks about.
The 10-second decision framework
If you're a marketer this quarter, here's how to choose:
Mostly briefs, reports, decks? → Cowork
Building reusable systems? → Code
Heavy or long content runs? → Code
Quick angle or hook brainstorm? → Chat
Sanity-check before sending? → Chat
Persistent project agents? → Code
The rule of thumb: Chat for thinking. Cowork for doing. Code for compounding.
Step two: Encode your strategy as a Skill
This is the lever that changes everything.
A skill is a reusable playbook — a file that teaches Claude how your business does a specific kind of work. Your ICPs, your voice, your no-go zones, your frameworks. One file. Reused everywhere.
What goes into a marketing strategy skill?
Eight components:
Company overview — business model, stage, GTM motion, pricing, industry, audience
Tiered ICPs — not just audiences, but tiers ranked by maturity: scaling, building, first customer, anti-tier
Marketing advantages — if you look back a year and you've won the market, why? Founder presence? Proprietary data? Ecosystem?
Perceptions — the 3–4 key narratives you're driving from the customer's point of view
Positioning — what is the product, who is it for, what's the alternative, why is yours better
Four ways to grow — more of the same audience, new audiences, better efficiency, more value per customer
Big bets — these come from your advantages crossed with your growth levers
Goals — whatever stage you're at, the numbers go in
Once this skill is loaded, every request becomes:
"Use my marketing strategy skill to write this brief."
Strategy stops being a doc. It becomes the rails.
Step three: Write a tiny CLAUDE.md
The CLAUDE.md file is the master instruction file Claude reads on every session. It's the bridge between your strategy and your daily work.
What goes in a good one:
Who you are — role, company, one-line description
What you sell — offers, price points, audience
Your voice — do's and don'ts, with 2–3 example phrases you'd actually use
Hard rules — things Claude must never do, in bullet form
Default behaviour — "When I ask for X, do Y"
Working style — brevity, format, when to ask vs. assume
Reference files — pointers to your strategy skill, brand voice, past content
The whole file should be under 200 lines. You can build it in 10 minutes. It survives every session. And it turns Claude from a generic assistant into a team member who already knows your business.
Step four: Run a smoke test
Before you trust the system, stress-test it.
The smoke test is simple: "Brief this campaign for ICP A."
If the output nails your voice, references your positioning, and avoids your no-go zones — the system works. If it doesn't, you know exactly what to fix. Refine the skill. Tighten the CLAUDE.md.
Run it again.
This is how the system works. Every correction you make gets written permanently into the skill file. The next 50 runs are different because of one correction.
That's the difference between a tool and an operating system.
Conclusion
AI-native marketing is about building a system where strategy no longer lives in a deck nobody opens. Where every brief, every campaign, and every prioritisation call runs through the same filter — your filter, encoded once, applied everywhere.
The horizontal AI players can't customise that filter for you.
That's the marketing advantage the rest of us actually have.
The question to ask your team this week:
What's the last decision we made that we couldn't trace back to strategy?
If the answer makes you uncomfortable, that's your starting point.

Comments