By Paul Stefanski in Systems Thinking — 25 Aug 2026

The File That Stops You From Re-Briefing Claude Every Monday

A CLAUDE.md file is a plain text file, saved at the top level of a project folder, that Claude reads automatically at the start of every session, before anyone types a word.

Claude.md

You open a new Claude Code session Monday morning to keep building the intake tracker your assembling for your PMO. First message: who you are, what the project's for, the naming conventions your team uses, and the one rule you repeats most: status updates get written in outcomes, not activity. Same briefing Tuesday. Same again Thursday, after adding two fields to the tracker and forgetting, again, to mention the naming rule until Claude got it wrong.

Nothing in that exchange is a mistake, exactly. It's what happens when every new session gets treated like a new hire's first day, every single time. Here's the part that should stop you: the retyping is hers, every Monday, on a rule she already gave Claude twice before.

Why the Default Approach Fails

Most people treat every Claude session as a blank conversation. Say what's needed, once, into the chat window. Get an answer. Move on. It looks efficient in the moment, and for a single task it usually is. The trouble starts on the second session, and the tenth, when the same standing facts have to be retyped because a new session carries nothing over from the last one.

That's a structural fact about how a session starts: with nothing remembered. The real cost shows up later, as correction fatigue. Retyping a naming rule once feels reasonable. Retyping it every session, indefinitely, means eventually it stops getting retyped, and the mistake comes back. Multiply that across every standing fact a project depends on, and the chat window becomes the only place any of it lives.

A session that starts with zero memory needs somewhere for that memory to live. That's the whole idea behind a CLAUDE.md file.

What a CLAUDE.md File Actually Is

A CLAUDE.md file is a plain text file, saved at the top level of a project folder, that Claude reads automatically at the start of every session, before anyone types a word. No special format required: headers with #, bullet points with -, and nothing else to learn.

Same project, two ways of starting Monday:

Without it: You type your project context, naming conventions, and the outcomes-not-activity rule from memory, hoping you does not forget one.

With it: You wrote that context once, saved it as CLAUDE.md, and Claude opens the session already knowing it.

Same project. Same AI. Two completely different starting points.

Writing One, In Practice

Creating it manually: In your project folder, create a new file named CLAUDE.md using any plain text editor — no special tool required. Start with a few headings that match how your project is organized, then fill in only what Claude cannot infer on its own: standing preferences, naming rules, structural facts about how the work is organized ("always write status updates in outcomes, not activity"). A blank file with clear structure is better than waiting for the perfect starting point.

Deciding what belongs: Only things worth repeating every session earn a place in the file. A detail that matters for one task, once, belongs in that session's chat rather than in the file. If the same correction shows up twice, that is the signal to move it into CLAUDE.md. Standing instructions, naming conventions, and persistent context about how work is organized are all fair game.

Keeping it usable: A short file gets followed; a long one gets skimmed. Aim under roughly 200 lines, and be specific rather than vague ("use 2-space indentation" works, "format things properly" does not). If two instructions end up contradicting each other, Claude may just pick one at random — so a periodic re-read to catch conflicts is worth the few minutes it takes.

What Skipping It Costs

Back to you. Six weeks in, without a CLAUDE.md, you had stated your outcomes-not-activity rule at the start of a dozen-plus sessions, and Claude has still gotten it wrong twice. Both times the instruction had scrolled far enough back in a long session that it stopped feeling worth repeating. With a CLAUDE.md, that rule gets stated once. It is the same instruction on session forty as it was on session one.

The retyped minutes add up across weeks, but the sharper cost is consistency. Without a standing file, how the project gets handled depends on whether that day's session happened to include the right reminder. With one, it depends on nothing but opening the folder. That's the real trade: a few minutes writing the file once, against however many sessions the project runs for.

Task-trackers keep re-explaining a project from memory, session after session. Strategic PMs write the outcomes and the assumptions down once, the same discipline behind the Four Pillars of Project Clarity, and let the file carry it from there.