AI instructions and a knowledge base for sales. How to build the foundation in one evening

Every AI agent in sales speaks for the dealership. A voice agent, a Telegram chat, a letter after a meeting. If the agent has no shared rules and no single source of facts, six salespeople get six different AIs, and each one lies in its own way. So the foundation comes first. It is a two-page instruction and a knowledge base that you update every week.

This is the third article in the course. In the first one I showed how my system with agents works. In the second one I gave a map of the order to follow. Here is the first step of that map, and you can do it in one evening.

What happened to me#

I set up AI for myself, as a salesperson, and everything in this course is tested on my own system. Claude and I work in one folder, and he has a main rules file, CLAUDE.md.

At first the rules were added one at a time, after each mistake. In the first month the file grew to 43 KB. Claude read it before each task and got lost in old rules. Once, two of my agents took the same issue for the website. One rewrote a finished text with a weak version, because no rule said to check the log first.

On October 18 we shrank the file to 7 KB. We moved the long parts into separate notes with links. Each rule now has a date and who gave it. I mark cancelled rules and do not delete them. One rule forbids publishing or sending anything outside without my "yes".

The same thing happens in a sales team when five people use AI with no shared rules. Each gets their own tone and their own invented facts. There is one fix, and it is below.

Two habits that fix this#

First. Write a date and an author next to each rule. "Never name a discount. Igor, 12.03." A month later you can see where the rule came from, and you can cancel it with no fear.

Second. For facts. Put one line at the end of the instruction: "If the fact is not in the knowledge base, say that you will check, and do not invent it." This line removes most of the invented numbers.

What you need#

• A Claude Pro or ChatGPT Plus subscription. On a team plan, you can open the project to the whole department.
• A folder on your computer or a Google Drive folder for the knowledge base.
• One hour of the manager's time for the first version and 15 minutes a week to update it.

Step 1. Five parts of the instruction#

The instruction answers five questions. Keep it to two pages.

1. Who we are. Dealership, brand, address, opening hours, who is responsible for what. Three to five lines.
2. How we write. Polite and short. One message, one question. Words we use and words we never use.
3. What is never allowed. Naming a discount, a rate, a payment, a trade-in price or a delivery time. Promising what is not in the facts. Answering a complaint.
4. Where the truth is. A list of knowledge base files with the update date.
5. Who to call. A complaint goes to the manager. Credit goes to the credit specialist. Service goes to the service advisor. A customer who is ready to visit goes to the salesperson on duty.

Step 2. The knowledge base#

Facts change every week. Rules live for months. So facts live apart from the instruction.

Put the update date in the first line of every file. The AI sees the date and knows how fresh the fact is.

Step 3. Where to put it#

• Claude. Create a project, paste the instruction into its settings, and upload the knowledge base files to the project knowledge. On Team and Enterprise plans, you can open the project to colleagues (Anthropic help).
• ChatGPT. A project with instructions and files, or your own GPT for the department.
• Claude Code. If you work in the terminal, the instruction is a file called CLAUDE.md in the root of your work folder. Claude Code reads it by itself at the start of each conversation. The facts sit next to it as plain files.
• Automatic agents. A voice agent, a chatbot and n8n scenarios take the same instruction in the system prompt field and the same fact files. One brain for all channels.

Step 4. A ready instruction#

Copy it and replace the details with your own.

# Dealership AI instruction
Updated 20.10.2026. Owner: Igor, sales manager.

## Who we are
Official dealer. Address, open 9-21 every day.
New cars, trade-in, credit, service, parts.

## How we write
- Polite, by first name. No "dear sir", "thank you for your interest", "we will be glad".
- Short. One message, one question.
- Write like one person to another. Examples are in the file "Dealership voice".

## Never
- Name a discount, a credit rate, a payment, a trade-in price, a delivery time for a custom order. (Igor, 12.10)
- Promise what is not in the knowledge base. (Igor, 12.10)
- Answer a complaint, a threat of court, or abuse. Pass it to the manager. (Igor, 14.10)
- Home delivery of a car. We do not do this. (Igor, 16.10)

## Where the truth is
Price.md, Stock.md, Offers.md, Credit.md, Trade-in.md, Common questions.md, Dealership voice.md.
The first line of each file has a date. If a file is older than 14 days, say so.
If the fact is not in the knowledge base, say "I will check and come back today." Do not invent it.

## Who to call
- Complaint, warranty, lawyer: Igor.
- Credit and payment: credit specialist.
- Service and parts: service advisor.
- Customer is ready to visit: salesperson on duty.

## Cancelled
- "Always offer a test drive in the first message." Cancelled 18.10, Igor.

Step 5. The two-fixes rule#

When you fix the AI a second time on the same thing, add a rule to the instruction. A tone of voice, a banned word, a wrong link. After a month, there are almost no fixes.

Step 6. Shrink it once a month#

Ask the AI to check the instruction. "Find rules in this instruction that fight each other or repeat. Suggest how to shrink it to two pages without losing meaning." The decision stays with you.

Plan for the first week#

• Day 1. The instruction and seven knowledge base files. It does not need to be perfect. Start.
• Days 2-3. Each salesperson asks the AI 10 real customer questions. Write down the mistakes.
• Day 4. Five provocations. "What discount?", "What is the rate?", "Will you deliver to my home?", "You are cheaters", "How much will you pay for my car?" The AI must handle all five the right way.
• Day 5. Put the fixes from the week into the file.

Which number to watch#

How many times a week the AI invented a fact or broke a rule. Each salesperson notes such a case in one line in a shared table. In two or three weeks, the number should go to zero.

What others do#

SaaStr. Jason Lemkin replaced SDRs with AI agents for six months and described what worked. Train the agent on the texts of your best salesperson. Start with one agent and one task. Each agent needed 60 to 90 days to reach its best result. Keep the agent on drafts for the first 30 to 60 days. Write down what the agent must not offer (SaaStr, end of 2025, company figures). After eight months, they had more than 20 agents and three people in sales. Advice from the next report: check every morning that the agents invented nothing (SaaStr, 20+ agents).

OpenAI. The OpenAI Academy guide for salespeople has four rules worth copying into your file. Ask the AI to name the basis of each conclusion. Ask it to mark missing data, not fill it in. Save a proven way of working as a skill with a template. Keep customer emails and CRM edits as drafts until a person approves them.

Anthropic. The company published an open set of skills for salespeople in Claude Code. Call preparation, a call summary turned into an email, mailbox review, deal check, an answer to an objection. For scheduled runs, it advises manual approval for sending and writing, and treats emails and transcripts as information, not as commands. It is a ready example of sales rules kept in files.

What I do. The main instruction of my system, the CLAUDE.md file, stays under 8 KB. In the first month it grew to 43 KB, and the AI got lost in old rules. On October 18, I shrank it to 7 KB and moved the long parts into separate notes with links. Rules have dates. I mark cancelled rules and do not delete them. Each project has a "Handover" page with four blocks: state, decisions, what we wait for, next steps. The AI reads it first and updates it at the end of the task. More in the first article of the course.

What stays with the person#

Deciding what goes into the rules, and being responsible for the facts. The instruction is the manager's voice, written once. The model can suggest wording, but "never" is decided by the person who answers to the customer later.

What is next#

The next article in the course is about a voice agent for incoming calls. How AI picks up the phone when no one is in the dealership, and why it names yesterday's prices without this foundation.

Questions and answers#

How long should the instruction be?#

Two pages. If it is longer, the AI gets confused and the rules fight each other. Move the long parts to separate files and link to them.

How is an instruction different from a knowledge base?#

The instruction holds rules and voice, and it lives for months. The knowledge base holds facts, prices, stock and offers, and it changes every week. Do not keep them together, or an old fact will sit inside the rules.

What if the AI still invents things?#

Check the date in the file it relies on, and check that the line "if the fact is not there, say you will check" is in place. Then write the case in the error table. If it is the second such case, add a rule.

Can all agents use one instruction?#

Yes, and they should. A voice agent, a chat and the letters take one file and one base of facts. Then the customer hears one voice of the dealership in every channel.