Agent Foundations for Marketers · Lesson 4 of 8

AI context file: what it is and what it changes

One file the agent reads every time, holding your positioning, your voice, and your hard-won constraints. It is what stops the output sounding generic.

The page you built last lesson read competent and a little bland, the kind of thing that could belong to any company in your category. That is not a weakness in the agent. It knows everything about marketing in general and nothing about you in particular. Closing that gap is the single most useful thing in the whole course.

Feel the difference first

I keep a practice company for this: Tempo, a fictional lifecycle-marketing SaaS. Its folder holds 298 customer reviews, eight sales call transcripts, usage and billing data, competitive briefs, and its own published positioning. No pages to design, no product needed. The material is the shape of marketing work wherever you do it.

Assignment: fetch the practice company

Paste this in the folder you already have open; nothing about your setup changes. Approve the commands it asks about, the same as during the build, and give it a minute to pull everything in. If it pauses to ask whether the download is safe, good: there is nothing harmful in this folder. Asking the agent to read through what it fetched before you trust it is a habit worth keeping.

Download the practice company from github.com/nicolaas-spijker/pmm-agent-demo into a folder called tempo inside this project, as a plain archive copy rather than a git clone. If it contains a workshop folder, delete it; that material is for a live event, not for this course. When it is in, list what is inside and stop.

Now ask the question a chat could never answer.

Assignment: test the positioning against the reviews

You are asking the agent to read a whole company and check its public story against what its customers say. Reading everything can take it ten minutes. That is the work happening, not the agent hanging, so let it run.

Read everything in the tempo folder, starting with its CLAUDE.md. Then tell me: what are the biggest complaint themes in the reviews, and does the positioning in tempo/positioning/live-messaging.md survive them?

What comes back is specific. It names themes with counts behind them, quotes the reviews, and tells you where the published positioning contradicts what customers actually say. Nothing about the agent changed between last lesson and this one. The folder did.

A ten-minute read before every task is no way to work, though. The job of making the context stick falls to one short file: Tempo’s CLAUDE.md, the one the prompt had the agent read first.

The file the agent always reads

Every one of these tools has a file it loads at the start of each session, before you have typed a word. In Claude Code it is called CLAUDE.md. Codex reads AGENTS.md. Gemini’s editor uses GEMINI.md, and Cursor keeps a rules file. Same idea everywhere: a plain text file, sitting in your folder, that the agent treats as standing instructions.

Definition · .md file

A file ending in .md is a plain text file. The md stands for Markdown, a way of adding structure with ordinary characters: a # starts a heading, a - starts a list item. Any editor can open one, no special software involved, and agents read and write them natively. Nearly every file in this course, the context file included, is a .md file.

You put in it the things you are tired of re-explaining. Every draft, every page, every reply then starts from your context instead of from the generic middle of the internet. The file applies itself.

Inside Tempo’s folder

The pattern that holds up is a short hub file plus a folder of depth. The hub gets read first, every time; the depth gets read only when a task calls for it. Tempo, sitting in your project right now, is built exactly this way. Hover or tap the marks for what each piece holds:

tempo/
The hubDepth, on demandFor the last lesson
├── CLAUDE.mdTempo’s hub: the standing facts up front, with pointers to every source below.
├── README.mdThe front door for humans browsing the repo. The agent’s briefing is the hub.
├── company/Ground truth the hub points to first: what Tempo is, who buys it, the voice, and the competitive set.
├── data/The 298 reviews plus usage and billing exports. The numbers behind every claim.
├── sales-calls/Eight call transcripts: objections, competitors, and why deals were won or lost.
├── competitive-intel/Four quarterly competitive briefs. Diffing them shows what is rising, not just what is.
├── positioning/The published messaging and pricing. Deliberately stale: claims to audit, not truth to trust.
├── connectors.mdEach export mapped to the live tool it stands in for: reviews for G2, calls for Gong, and so on.
├── output/Where finished briefs land. Empty until the agent writes one.
└── .claude/commands/Three saved workflows. The last lesson has you run one and read it from the inside.

What goes in the hub file

Keep it lean. Anthropic’s own guidance for Claude Code puts a number on it: aim under 200 lines, because long always-loaded files dilute how well the agent follows any single line. Four short sections cover it.

  • Positioning. What you are, who it is for, and the one or two things you do that the alternatives do not.
  • Voice. How you sound, with a line or two of do and do not. “Plain and direct, no hype, no emdashes” does more than a page of adjectives.
  • Customer. Who you are talking to and what they already know, so the agent stops explaining things they understand.
  • Constraints. The decisions you learned the hard way. “We do not discount.” “Never call us an agency.” The lines you would correct a new hire on twice.

Where everything else goes

As a marketer you sit on far more intelligence than 200 lines. Personas, battlecards, win/loss notes, research decks.

  • Needed on every task, sayable in a line or two: the hub.
  • Needed on some tasks, or needs real depth: its own file in docs/, with one pointer line in the hub.
  • Everything else: nowhere.

The docs/ files can afford to be long. One topic per file, and a marketer’s docs/ fills out to something like this:

docs/
All read on demand
├── personas.mdThe full personas and ICP detail.
├── messaging.mdThe messaging framework.
├── battlecards/One battlecard per competitor.
├── win-loss.mdWin/loss notes.
├── pricing.mdThe pricing rationale. The settled one-line rule stays in the hub.
├── quotes.mdThe customer-quote ledger.
├── research/Summaries of your research reports. The originals live in an archive, not here.
└── roadmap.mdThe roadmap, dated, since it moves.

Tempo keeps its depth folders at the root instead of tucked into docs/; either shape works.

And some intelligence genuinely belongs nowhere.

  • Live numbers. Pasted metrics go stale within weeks, so have the agent pull them from the source at task time.

  • Generic marketing wisdom. Already in the model.

  • Anything the agent can read for itself in the folder.

  • Raw transcripts and report dumps. Summarize into docs/, archive the original.

One test settles most edge cases: if you would not repeat it to a capable freelancer in their first week, it does not belong in the always-loaded file.

Skills plug into the same hub

Skills are the second layer: repeat routines saved so they run on one instruction. A skill is a small folder with its own instructions file: a workflow described once, plus a line saying when to use it. That line is the trigger. The agent reads it and reaches for the skill on its own whenever a task matches.

Skills only earn their keep on top of this file. A skill says how to run a job; the context file makes it come out sounding like you. A hundred borrowed skills without it produce someone else’s marketing, efficiently.

I make the longer case against stacking skills in the Build Log. One honest file first. Then a handful of skills you wrote yourself.

Assignment: read the file the agent reads

Tempo’s file is the grown version of the shape you just learned: a hub that keeps the standing facts up front and points into deeper files for the rest. The agent can give you a guided tour.

Open tempo/CLAUDE.md and walk me through it: what each section tells you, which lines change how you would write for this company, and one thing you would still want to know that the file does not say.

Every prompt here pointed at Tempo’s file by name, because tempo/ sits inside your project rather than at its root. A context file loads by itself only when it carries your tool’s filename and sits at the root of the project you opened.

Writing your own gets its own sitting in Build your marketing brain: the interview that writes your file, your data wired in, and your first skills saved. What comes out will be a first draft on purpose, because real work keeps teaching a file like this what belongs in it.