The first real automation I built in Claude Code kept making a decision that was subtly wrong for our business. It scored leads like we sold to enterprise when we sold mid-market, and it did it confidently, every run. I kept correcting it in the chat, run after run, like an idiot with a mop instead of a plumber.
The problem was not the agent. I had never written it a CLAUDE.md, so it was guessing at who we are, and guessing wrong. CLAUDE.md is the standing brief the agent reads before it does anything, and it is where that context is supposed to live.
It is the single highest-value file in a Claude Code setup, and the one operators skip. Here is what goes in it and why it fixes the guessing.
CLAUDE.md is the agent's standing brief
When Claude Code starts, it reads CLAUDE.md. Think of it as the onboarding doc you would hand a sharp new hire: how things work here, the rules, and the context they need to make good calls without asking you every time. You can keep one at the project level and one at your user level for defaults across projects.
Everything else in this pillar leans on this file. A well-written CLAUDE.md is the reason the agent's output fits your world instead of a generic one, and a missing one is the reason it keeps making plausible, wrong decisions.
TipWrite CLAUDE.md the way you would brief a new hire on day one. If you would say it out loud to someone starting Monday, it belongs in the file.
What to put in it
Three layers earn their place: the project facts (how it is structured, how to run things, the key files), your rules (conventions, do's and don'ts), and your domain context (your ICP, your positioning, what you sell and to whom). The third is the one GTM operators forget, and it is the one that made my lead-scoring agent stop guessing.
Be concrete. 'We sell to mid-market RevOps leaders, not enterprise IT' is a fact the agent can act on. 'Understand our customers' is a wish it cannot. The specificity is what turns a standing brief into standing behavior.
Keep it tight
A CLAUDE.md is not a wiki. The longer and vaguer it gets, the less of it the agent actually honors, and a bloated file crowds the context window that the real task needs. Write the rules that matter, cut the rest, and treat length as a cost.
The way it earns its keep over time is through edits, not size. When the agent repeats a mistake, do not re-correct it in the chat, add one line here. That is the whole discipline: turn a correction you keep making into a rule you make once.
TipWhen you find yourself correcting Claude Code on the same thing twice, stop and add the rule to CLAUDE.md. The third correction should never be necessary.
Mind the context budget
Everything the agent reads, CLAUDE.md, the files it opens, the running conversation, competes for the same context window from the concepts. A focused CLAUDE.md leaves room for the task; a sprawling one spends the budget on itself. When a long session starts drifting, that is usually the window filling up, not the model getting worse.
So treat CLAUDE.md as high-value, low-volume: the few things that must always be true, stated once. Depth of context comes from the right lines, not the most lines.
Where CLAUDE.md goes wrong
The first failure is not having one, which is my scar: the agent guesses your business and guesses wrong, confidently, forever. The second is letting it rot, keeping positioning or an ICP that changed months ago, so the agent is now confidently out of date, which is worse than a gap because it looks authoritative.
The third is the wiki problem: stuffing it with everything until the agent skims it and honors none of it. All three have the same fix, treat it as a living, tight, curated brief, not a dumping ground and not an afterthought.
TipReread your CLAUDE.md when your positioning or ICP changes. A standing brief that is quietly stale steers every run the wrong way.
The GTM version
For a go-to-market setup, CLAUDE.md is where your ICP, positioning, voice, and rules live, so every brief, email, and analysis the agent produces is grounded in your actual business. It is the file that made the difference between my agent guessing and my agent knowing, and it is why the site's automations assume it exists.
Write it once and every run afterward inherits it, which is the same compounding you get from a Project in the apps, here in the tool where you build the automations.
So before your next Claude Code run, ask: does it know who you sell to and how you talk, or is it about to guess? What is the one line that would have saved your last three corrections?
TipSeed CLAUDE.md with your ICP, positioning, and voice first. Those three lines fix the majority of generic, off-business output before you write anything else.
How to set it up
Create a CLAUDE.md in the project root
Claude Code reads a file called CLAUDE.md at the start of every session and treats it as standing orders. Make one in the folder you work from.
Put the standing brief in it, including your GTM context
Write down what you would tell a sharp new hire on day one: how the folder is organized, your conventions, and the domain context that never changes between tasks. For GTM that means your ICP, your positioning, and your rules. For example:
"This workspace runs our outbound. accounts.csv is the target list. We sell an inbound demand platform to VPs of Marketing at B2B SaaS from 50 to 500 employees. Voice: peer to peer, no hype words, lead with the number, emails under 120 words. Never invent a stat."
Keep it tight
Everything in CLAUDE.md is read into context on every run, so a bloated file quietly costs you room and money. Keep it to rules you could catch being broken. Prune it when something stops being true, the same way you would fix an onboarding doc that lies to new hires.
TipIf a line in your CLAUDE.md is a nice-to-have rather than a rule, cut it. Standing brief, not a wiki.
Frequently asked questions
What is CLAUDE.md?
A file Claude Code reads at the start of every session, holding the standing context and rules for the work: project facts, your conventions, and your domain context. It grounds the agent so you do not re-explain your setup, or watch it guess.
What should I put in CLAUDE.md?
Three things: project facts (structure, how to run it), rules (conventions and guardrails), and your domain context (ICP, positioning, what you sell). For GTM, the domain context is what stops the agent from guessing about your business.
Can I have more than one CLAUDE.md?
Yes. A project-level file holds context for that work, and a user-level file holds your defaults across projects. Keep each focused on what belongs at that level.
Why does Claude Code keep making the same mistake?
Usually because the rule it is breaking is not written down. Add the missing guidance to CLAUDE.md rather than correcting it every session; the file is where a repeated correction becomes a standing rule.
How long should CLAUDE.md be?
Short enough to be honored. A tight file of the rules that matter beats a long one the agent skims, and it leaves more of the context window for the task. Treat length as a cost.
What is the most important thing to put in it for GTM?
Your ICP, positioning, and voice. Those three fix the bulk of generic, off-business output, because they are exactly what the agent would otherwise guess, and guess wrong.
Does CLAUDE.md use up the context window?
Yes, everything the agent reads competes for the same window. That is why a focused CLAUDE.md matters: it grounds the agent without crowding out the task. High value, low volume.
How is CLAUDE.md different from a Project in the apps?
Same idea, different surface. A Project holds persistent context in the Claude apps; CLAUDE.md holds it in Claude Code. Both mean you set your context once and every run inherits it.
Sources & further reading
Claude ships fast. This page was last reviewed Aug 22, 2026; verify time-sensitive details against the official docs above before relying on them.