Why this matters
Good engineering teams put a README at the top of every repo so a new hire can get productive without asking twenty questions. Your business needs the same file. Each agent is a new hire: it starts out not knowing who you sell to, what you charge, or what you would never say to a customer.
Without a shared file, that context lives in your head and gets retyped into every prompt, a little differently each time. That is how a support agent ends up promising a refund policy you do not have. A business README gives every agent one source of truth. When something changes, you update it in one place.
What goes in it
Separate durable business context from the details that change month to month.
| StableReview quarterly | |
|---|---|
| What we sell | What it is, who it is for, in the customer's words |
| The customer | Who buys, who uses it, and the exact words they use for the problem |
| Offer and pricing | Plans, prices, what is included, what is not, and refund or cancellation policy |
| How we are different | Who you lose to, and why people pick you anyway |
| Voice | Checkable writing rules drawn from real examples |
| Decisions log | Choices already made, with one line on why each |
| Hard rules | What an agent must never say or promise, and when it hands off to a human |
| CurrentReview monthly | |
| Right now | This month's priorities, the numbers you are watching, what is launching, and the biggest open problem |
| Glossary | Product names, internal terms, and acronyms |
The decisions log keeps agents from reviving choices you already made. Hard rules keep a customer-facing agent from promising something you cannot offer.
Build it with an interview, not a blank page
Run the prompt below in a fresh chat with whoever knows the business best, which is usually you. Keep your pricing page and policies open so you can give exact wording. It asks one question at a time and turns your answers into a file agents can use.
I want you to write the README for my business: the single file every AI agent I build will read before it does anything, whether it is answering customers, writing content, or writing code. It has to capture how this business actually works, not the pitch-deck version.
Interview me one question at a time. Wait for each answer before you ask the next. Cover these areas in order:
1. What we sell, to whom, and what problem it solves. Ask me to say it the way a customer would, not the way the website does.
2. The customer: who buys, who uses it if that is someone different, what they were doing right before they found us, and the exact words they use to describe the problem.
3. The offer: products or plans, prices, what is included, what is not included, refund and cancellation policy, and anything customers regularly get confused about.
4. How we are different: who we lose deals to, and why customers pick us anyway.
5. Decisions we have already made: tools, channels, pricing choices, and things we tried and stopped. Ask me why for each one.
6. Hard rules: things an agent must never say, promise, or do, and the situations where it must hand off to a human.
7. Right now: this month's priorities, the numbers I am watching, what is launching, and the biggest open problem.
CALIBRATION
After the interview, ask me for two things:
- The best customer-facing thing we have written, such as an email, product page, or support reply. Note what makes it work in specific terms.
- An AI draft for the business that I would never send. Note exactly what is wrong with it.
Use both to write voice rules. Every rule must be specific enough to check. "Use the customer's words for the problem, not ours" is a rule. "Be friendly" is not.
PUSHBACK RULES
- If I describe the business in marketing language, ask what a customer would actually say.
- If I am vague about pricing or policy, ask for the exact number or wording. Agents will repeat whatever is in this file to customers.
- If two answers contradict each other, point it out and ask which one is true.
- If I mention a product name, internal term, or acronym you do not know, ask before moving on.
- Stop after about 20 questions, and tell me when you are about to wrap up.
WHAT TO DELIVER
Create a file called BUSINESS.md in two parts.
PART 1: STABLE (review every quarter)
- What we sell and to whom, in three sentences using customer language
- The customer: who they are, their problem, and their words
- Offer and pricing: plans, prices, inclusions, exclusions, and policies
- How we are different, and who we compete with
- Voice rules drawn from the calibration pair
- Decisions log: each decision, with one line on why
- Hard rules: never say, never promise, always escalate when
PART 2: CURRENT (update every month)
- Last updated: today's date
- This month's priorities, in order
- The numbers that matter right now, and the target for each
- What is launching or changing
- The biggest open problem
- Glossary: product names, internal terms, and acronyms
RULES
- Keep the file under 900 words total. Agents read every word, every time.
- Use my exact numbers and policy wording. Do not round, soften, or fill gaps with guesses.
- Mark anything I was not sure about as [UNCONFIRMED] so no agent repeats it to a customer.
- End with up to five questions you would ask me in a second session to fill the gaps.
Where to put BUSINESS.md
- Coding agents: Save it in the repository root. Add one line to
CLAUDE.mdorAGENTS.md: “ReadBUSINESS.mdbefore any work that touches customers, pricing, copy, or product decisions.” - Claude or ChatGPT Projects: Add the file to the knowledge for every project that does business work, such as support, content, or sales.
- Agents you build: Load it into the system prompt or have the agent fetch it at the start of each run. One edit then updates every agent.
- Your team: Keep it somewhere people can edit, and date every change the same way you would treat code.
Do this next
Pick one agent or saved prompt you already use for the business, such as a support reply or a content draft. Rerun its last task with BUSINESS.md loaded and compare the two outputs. Every difference shows you something the old version had to guess.