claude.md for founders: the one file that stops context drift.

elisabeth hitz · june 26, 2026 · 6 min read

if you are a founder using Claude for real work, you have already felt context drift. you explain your customer on monday, and by thursday the model writes like a generic SaaS blog again. CLAUDE.md is the fix most people skip because it sounds technical. it is just a markdown file in your project root that Claude reads before every session.

think of it as the employee handbook for your AI. not a prompt you retype. a standing file that pins voice, rules, and what matters right now.

what goes in a founder-grade CLAUDE.md

a useful CLAUDE.md is short and opinionated. founders who treat it like a wiki get drift anyway. the file should answer four questions in plain language:

• who we sell to. one ICP paragraph, the pain in their words, what they already tried.
• how we sound. lowercase or title case, banned words, proof rules (no made-up stats).
• what we sell right now. active offers, prices only if they are public on the site, where buyers land after pay.
• what jobs AI runs this month. launch, distribution, nurture, offers, or brain setup. name the operator, not "help with marketing."

that last bullet is where most founders stall. they ask for "better copy" when the real job is launch sequencing or a distribution calendar. naming the job title changes what Claude optimizes for. that is the difference between a launch operator and another random prompt.

what to leave out

do not paste your whole notion workspace. do not store daily task lists that go stale. do not dump competitor research without a verdict. CLAUDE.md is not a database. it is the minimum context that keeps output on-brand when you are moving fast.

if a section needs more than 40 lines, it belongs in a skill file or a reference doc the skill points to. that is how prompts graduate into skills without bloating the root file.

how this connects to operators

CLAUDE.md is the floor. operators are the job titles on top. the five GTM operators (launch, distribution, nurture, offers, brain architect) each ship as a skill file with a checklist and output format. CLAUDE.md tells Claude who you are. the operator tells it what to produce today.

start with brain architect if every chat still opens cold. start with launch if you keep planning in circles. the free map at /operators/ tells you which one to install first based on where you are stuck.

a minimal starter template

copy this skeleton into CLAUDE.md and fill the brackets. keep it under one screen.

# project
## ICP
[one paragraph]
## voice
[3-5 rules]
## active offers
[name + URL, no prices unless public]
## this month's operator jobs
[launch / distribution / nurture / offers / brain]
## proof I can cite
[real numbers only]

update the "this month" section when your bottleneck moves. everything else should change slowly. that is how context stops drifting without another subscription.