The 4 .md Files That Turn Claude Code Into a Marketing Machine
Before you build a single skill in Claude Code, you need a foundation layer. These 4 files are that layer — and they power everything else.
The marketers who get the best results from Claude Code all have one thing in common: they build the foundation layer first. These 4 .md files are a great starting layer, and they’ll be able to power the marketing skill that sit on top of it.
When Pixar made Toy Story in 1995, they nearly didn’t finish it. The story kept breaking. Andrew Stanton, Pete Docter, and Joe Ranft would solve a narrative problem on one film, then watch a different team hit the exact same problem on the next one.
Every film starts from zero. The lessons lived in individual directors’ heads; there was no shared system.
Ed Catmull wanted to fix the problem. The problem wasn’t talent. He already had the best directors. His fix was instead a system called Braintrust—a shared creative intelligence layer where directors screened work-in-progress and drew on collective context across every film. It provided shared knowledge about what was working, what wasn’t, and why.
Did it work?
Disney acquired Pixar in 2006, and Catmull exported the Braintrust to Disney Animation as the “Story Trust.” Disney had the same animators, same budgets, and had been making bad film after bad film. After adding the shared context layer, those same people produced Frozen, Big Hero 6, Zootopia.
Same talent. New system. Better outcomes.
That’s the exact problem I see marketers hitting in Claude Code.
They build skills in LinkedIn post writing, newsletter drafting, and ad copy generation. Each skill is good. Each skill works. But each skill starts from zero every time. No shared understanding of the audience. No consistent voice. No awareness of the competitive landscape. No memory of what converts and what doesn’t.
The output is... fine. Generic. Interchangeable with any other SaaS marketer’s content. The same way Pixar’s directors were individually brilliant but collectively inefficient without shared context.
The fix is the same: build the foundation layer first.
What I Mean by “Foundation Layer”
If you’ve read my previous posts, you know I’ve built a marketing system in Claude Code — skills for researching, drafting, enriching, and reviewing content.
But the skills aren’t what make it work.
Underneath every skill sits a set of .md files that teach Claude my audience, my marketing style, my services, and products. Not CLAUDE.md (the system configuration file everyone’s already talking about). These are different. These are files the marketer creates, about their audience, their voice, their market, their buyer’s journey.
I call them foundation files. They live in a /foundation/ directory, and every skill reads from them before it executes.
The difference this makes is night and day.
Without foundation files, a positioning skill results in generic positioning and stuffy web copy that could be about any SaaS company. With them: the skill knows your audience uses “agentic workflows” not “AI tools,” that your voice is direct and anti-hype, and that your competitive positioning is against specific alternatives.
The 4 Foundation Files
At this point, I’ve built a robust foundational layer for my marketing system. But, there are 4 .md files in that layer, I believe, that would benefit any marketer or marketing team who wants to use Claude Code.
Here’s the important part of this system. Skills don’t load all foundational files every time. That would dilute context and confuse the output. Each skill scans the foundation folder, reads the file headers, and only loads what it needs for that specific task.
A LinkedIn post skill loads Audience Delight + Creator Style.
A product launch messaging skill loads Market Positioning + Customer Journey.
A nurture email sequence loads Audience Delight + Customer Journey + Creator Style.
Selective loading. Like imports in code — you don’t import *.
Here’s each file and why it exists.
1. Audience Delight Profile
What it replaces: The traditional ICP
Here’s the thing about ICPs: they’re built for humans reading slide decks. “Decision-maker at a 200-500 person SaaS company, VP level, cares about pipeline efficiency.” That’s useless for an AI agent writing an engaging YouTube transcript.
An Audience Delight Profile captures what makes your audience light up. What they share unprompted. What they complain about when your brand isn’t in the room. What language do they use, the actual words, not the corporate version?
The insight that changed how I built this file:
Your audience has a vocabulary. When they talk about bad AI output, they say “slop” and “garbage output” — not “inaccurate results.” When they talk about deploying AI, they say “ship it” and “running in prod”, they don’t say “implement the solution.”
Get the vocabulary wrong, and your content sounds like a vendor. Get it right, and it sounds like a peer.
Here’s what the vocabulary section looks like in a completed Audience Delight file (this one’s built for Notion):
When Notion users talk about their workspace, they say “my Notion,” “my setup,” “my system”, they don’t say “the platform” or “the solution.” When they talk about organization, it’s “second brain” and “single source of truth”, never “knowledge management system.” Flexibility is “blank canvas” and “build what you need”, not “configurable enterprise solution.” And when they talk about other tools, it’s “tool sprawl,” “context switching,” and “tab hell”, not “technology stack consolidation.”
Every skill that writes audience-facing content reads this file. The result: your ad copy uses the same language your audience uses in Reddit threads. Your emails name the frustrations in their words, not yours.
How to build yours: Feed Claude your Reddit threads, sales call transcripts, support tickets, and competitor comment sections. Ask it to replicate this profile from their language, not your marketing department’s language.
» Download in html format here
2. Creator Style
What it replaces: The brand voice guidelines PDF
Traditional style guides are aspirational. “Our voice is warm, professional, and empowering.” That’s vague enough to apply to anyone.
A Creator Style file is extracted empirically from your best-performing content. How you actually sound when the content works.
The insight that makes this file different:
It captures patterns, not adjectives. Instead of “we’re conversational” you get “short sentences lead, longer sentences follow to explain, fragments are welcome for punch, rhythm alternates staccato then flow.” Instead of “we’re empowering” you get “always lead with what the user can do, not what we built.”
The most powerful section is the comparison — actual phrases that sound like your brand vs. phrases that don’t. For Notion, “Your workspace should adapt to how you work, not the other way around” sounds like the brand. “Our platform provides a streamlined solution for team collaboration” doesn’t. “Build your own CRM. Or a wiki. Or a project tracker. Or all three” sounds like Notion. “Notion offers comprehensive project management capabilities” doesn’t.
When a skill reads this, it has concrete calibration. Patterns it can match against, not vibes it has to interpret.
How to build yours: Gather your 10 best-performing pieces. Feed them to Claude. Ask it to describe the voice, tone, and patterns it observes. Then validate — does the description match the winners but NOT the losers? That’s the test.
» Download in html format here
3. Market Positioning Map
What it replaces: The positioning slide deck from last quarter
Most companies treat positioning like a point-in-time exercise. An offsite. A quarterly review. A slide deck that gets referenced once and then forgotten.
A Market Positioning Map is a living .md file. It tracks what competitors are currently claiming, what territory you own vs. what’s contested, where the white space is, and what’s shifting in the market right now.
The insight that makes this file valuable:
It’s not just “who are our competitors.” It maps three territories:
Territory you own — positions competitors can’t credibly claim. For Notion, that’s “flexibility without complexity” and their community-driven ecosystem. No competitor has a community creating templates and YouTube content at that scale.
Territory that’s contested — multiple players claiming the same thing. “All-in-one workspace” is contested. “AI-powered” is contested. When everyone claims it, no one owns it.
Territory you’ve ceded — and that’s fine. Notion doesn’t compete with Asana on deep project management methodology. Knowing what you don’t claim is as important as knowing what you do.
Why this matters for skills: Any skill that involves launch messaging, competitive differentiation, or content that stakes a claim reads this file. It prevents your AI from positioning you in territory you don’t own, or missing territory you do.
How to build yours: Audit competitor homepages (what are they actually claiming right now?), read G2 reviews (how do customers compare you in their own words?), and review your win/loss data. Update monthly — markets move faster than quarterly decks.
» Download in html format here
4. Customer Journey Intelligence
What it replaces: The funnel diagram on a whiteboard
Traditional journey maps are static. They describe stages. They don’t capture the actual moments where people convert, stall, or leave.
A Customer Journey Intelligence file captures the real mechanics: how people find you, what questions they ask during evaluation, what proof points close deals, where they get stuck, and what triggers expansion.
The insight I keep coming back to:
Most journey maps focus on what you want the customer to do. This file focuses on what the customer actually does. Including where they stall.
For Notion, the #1 stall point is blank canvas paralysis — people sign up, see an empty workspace, and don’t know where to start. The #1 churn reason is team adoption failure, one champion builds a workspace, nobody else uses it.
If your nurture email skill doesn’t know about these stall points, it writes generic “here’s what you can do with Notion” emails. If it does know, it writes emails that directly address blank canvas anxiety with a template they can duplicate in 2 minutes. Better yet, if it knows your conversion triggers you can turn curiosity into action.
How to build yours: Pull from analytics (actual conversion paths and drop-off points), sales call recordings (what objections close and what objections kill), and customer interviews (”walk me through how you found us and decided to buy”).
» Download in html format here
How Skills Auto-Load the Right Foundation Files
Here’s the killer part:
Every foundation file has a header that declares when it should be loaded:
Load this file when the skill: writes any content — blog posts, social copy, emails, landing pages, ad copy, or any output that produces words.
Do NOT load this file when the skill: only needs audience data, only needs competitive positioning, or only needs the buyer’s path.
And for every skill you need to add this one generic instruction block:
Step 1: Load Foundation Context. Scan ./foundation/ for all .md files. For each file, read ONLY the header block (everything above the first ---). Check the “Load this file when” line against this skill’s task. Load the full file ONLY for files where the criteria match. Skip all others.
That’s it. The intelligence lives in the files, not in the skills. You never have to manually wire dependencies. Build a new skill, paste the same 6-line block, and the foundation files route themselves.
Add a 5th foundation file next month? Every existing skill automatically picks it up if it’s relevant.
This compounds. Every update improves every skill that reads that file. You don’t improve one skill at a time. You're improving the entire system at the foundation level. That’s how you start to turn Claude Code into a marketing powerhouse.
Catmull's insight wasn't "hire better directors." It was "give them shared context and let them create." - give then a Braintrust.
Until Next Time,
Happy AI’fying
Kieran
This is a clever workflow, really appreciate the detailed walkthrough.
Same talent with shared context changes everything, and that's exactly what these foundation files do for AI output.