The One File that Controls How Smart Claude Is

I run delivery on GenAI and AR/VR programs across 120+ markets with over 60 team members.

Few of my projects nowadays run on me remembering every decision. They run on a singular file.

Most people don't know this file exists. Fewer still have set one up.

If you've ever opened a new Claude conversation and spent the first 10 minutes re-explaining a project you already explained last week, this is the fix: a markdown file.

What is a markdown file, and why does Claude read it automatically?

A markdown file is just a text file. It's not special software, in no proprietary format. What makes it different from a regular note is that Claude (and most other chatbots) will automatically read it at the start of a conversation, before it acts on anything you ask.

Put instructions in it — how you like status updates written, what "done" means on this project, who the stakeholders are, what's already been decided — and Claude shows up already knowing it.

No markdown file, and you're re-explaining the same context every time you open a new chat.

Caution: this only works if the file actually gets read. If you're pasting instructions into the chat window instead of saving them where Claude looks automatically, you're doing the re-explaining manually and just calling it something else.

Choose one project you're tired of re-explaining

Don't start with your whole job. Start with the one project where you've caught yourself typing the same background paragraph into Claude more than twice.

Open a conversation and attach any files that are already relevant — a scope doc, a stakeholder list, a status template. Or start with nothing and go straight to the interview in the next step.

What this looks like in a real workday:

• A recurring client program where the stakeholder list and reporting cadence never change, but you type them out fresh every week

• A cross-functional build where three teams have different definitions of "on track," and you keep re-stating whose definition wins

• A status report you rebuild from memory each Friday instead of from a saved standard

Caution: picking a project that's still in early scoping, where the goals themselves are moving, will produce a file that's out of date in a week. Pick something with a stable shape, even if the details inside it are still active.

Let Claude interview you to build the file

You don't have to write this cold. Paste a prompt asking Claude to interview you about the project — what success looks like, who's involved, what decisions are already locked, how you want updates formatted — and let it ask the questions.

Example prompt: "Interview me about [project name] so you can build a CLAUDE.md file for it. Ask me about scope, stakeholders, what's already decided, and how I want status updates written. Then draft the file from my answers."

Claude will come back with follow-up questions before it drafts anything. Answer plainly — the file is only as useful as the specificity you give it.

Caution: don't rush the interview to get to the file faster. A file built from three rushed answers reads generic, and a generic file is barely better than no file at all.

Keep the file under two hundred lines

This is the step most people get wrong, and it's the one that actually determines whether the file helps or hurts.

The instinct is to add more — more history, more edge cases, more detail — on the theory that more context means better output. It doesn't work that way, for two reasons:

• The longer the file, the longer Claude takes to read it before it can start on your actual request

• An overly specific file over-constrains Claude's thinking — you'll get answers that follow your instructions to the letter and miss the better option you didn't think to write down

Treat it the way you'd treat onboarding a strong hire: give them the context they need to operate with judgment, not a script that removes the judgment entirely.

Caution: if your draft comes back over 200 lines, don't just delete sections at random. Cut the parts that describe history — decisions Claude doesn't need to re-litigate — and keep the parts that describe how you want it to operate going forward.

Upload it to your project's context folder

Once you have the file, upload it into the project's context folder rather than pasting it into a single chat. This is what makes it persistent — every new conversation in that project reads it automatically, instead of only the one conversation you pasted it into.

After it's uploaded, Claude stops guessing at your context and starts operating inside it. Skip this step, and every gain from the interview and the drafting disappears the moment you close the chat.

Caution: if you update how the project runs — a new stakeholder, a changed cadence — update the file too. An out-of-date CLAUDE.md is worse than no file, because it actively steers Claude toward the wrong assumptions instead of just lacking them.