All articles

AI agents & Codex work

Stop Re-Explaining Your Business to AI: Build Its Operating Manual.

By · Updated June 2026

Short answer

Stop pasting your business background into every new chat. An instruction file like AGENTS.md acts as a permanent operating manual that AI agents load automatically. By defining your business context, offers, tone, rules, and escalation boundaries on a single markdown page, you turn repeated explanations into durable company infrastructure.

Stop Re-Explaining Your Business to AI: Build Its Operating Manual.

If every new AI session starts with “Here is what my business does…”, you do not have an AI system. You have a very fast intern with amnesia.

The biggest productivity loss with AI in small businesses is not picking the wrong model. It is the friction of repeatedly typing the same background context, customer profiles, quality standards, and negative constraints into empty chat boxes.

The mistake is treating AI as a disposable chat interface. The right first move is building an operating manual: a single instruction file (commonly named AGENTS.md or CLAUDE.md) that AI agents read automatically before doing any work.

Prompts vs Templates vs Operating Manuals

To understand why your current workflow feels chaotic, look at the three levels of AI guidance:

  1. One-Off Prompts: Ad-hoc questions typed into a chat. Useful for quick brainstorming, but zero institutional memory.
  2. Reusable Prompt Templates: Saved snippets you copy and paste. Better, but still requires manual effort and degrades when team members forget to paste them.
  3. The Operating Manual (AGENTS.md): A permanent, structured document stored directly in your workspace or project repository. Every agent reads it automatically at session startup, applying your company standards to every response.

The filename varies by tool, but the operating manual does not.


The 7 Sections of a Working AI Brief

A high-leverage instruction file does not explain everything about your business. It sets boundaries, defines sources of truth, and clarifies escalation paths across seven core sections:

# AGENTS.md - Company Operating Manual

## 1. Business & Audience
Who we are: [e.g., Regional residential solar installer in Catalonia].
Primary audience: [Homeowners with single-family houses looking to lower electric bills].

## 2. Offers & Core Services
- Turnkey photovoltaic installations (3 kW to 10 kW).
- Battery backup storage systems.
- Administrative management of local subsidies and grid feed-in compensation.

## 3. Voice & Tone Rules
- Direct, practical, jargon-free language.
- Speak peer-to-peer. Never use corporate buzzwords or inflated marketing hype.
- Never use em dashes. Use commas, parentheses, colons, or periods.

## 4. Rules & Negative Constraints
- Never promise specific subsidy approval timelines or guarantee financial grants.
- Never invent client names, case studies, or unverified statistical percentages.
- Always state assumptions clearly before performing technical calculations.

## 5. Sources of Truth
- Pricing and component catalog: `docs/pricing-grid-2026.md`
- Installation standard operating procedures: `docs/sop-solar-survey.md`

## 6. What You May Do Autonomously
- Draft email follow-ups based on real proposal notes.
- Format and structure technical summaries from supplier PDFs.
- Clean and normalize customer inquiry data.

## 7. What You Must Escalate to a Human
- Any request for custom pricing discounts outside our standard tier.
- Formal customer complaints or legal disputes.
- Final proposal sign-off before sending to the client.

Real Example: A Solar Installer’s Operating Manual

Consider what happens when an AI agent drafts a follow-up email without an operating manual versus with one:

  • Without AGENTS.md: The AI writes a generic, three-paragraph email filled with generic adjectives (“We are thrilled to present our cutting-edge solar solution!”), promises a 50% regional subsidy that expired last year, and uses an overly formal tone.
  • With AGENTS.md: The AI checks the source of truth, notes that subsidies require official administrative validation, adopts a direct tone, refers to the homeowner’s specific 5.4 kWp roof sizing, and leaves the final pricing confirmation for the sales manager.

Connecting Your Brief to SOPs and a Second Brain

Your AGENTS.md file is the gateway, not the entire library.

When your business knowledge lives across messy email chains, Google Docs, and employee heads, AI will always deliver generic output. By maintaining a structured company Second Brain in markdown tools like Obsidian, your instruction file can point agents directly to specific Standard Operating Procedures (SOPs) for repeatable execution.

To understand why clear operational steps must precede automation, read our guide on why SOPs come before automation.


The 45-Minute Setup Sprint

You can write your first complete operating manual today in under an hour:

  1. Minutes 0 to 15 (Drafting Core Facts): Fill in sections 1, 2, and 3 using the template above. State your business model, target client, and preferred communication style.
  2. Minutes 15 to 30 (Setting Constraints and Escalations): List the three worst mistakes an assistant could make with a client. Add them to Section 4 (Constraints) and Section 7 (Escalation).
  3. Minutes 30 to 45 (Linking Sources): Identify your two most important reference files (pricing sheet, service list) and list them in Section 5.

Save this file as AGENTS.md in your project folder, or copy it into your workspace project instructions. Your team now has a consistent, permanent baseline for every AI session.


If your team has this problem, this is what we work on in a practical AI training for small businesses. We help you build company-wide instruction files, structure your internal knowledge base, and deploy custom agent workflows that eliminate repetitive busywork.

Frequently asked questions

What is the difference between a prompt and an instruction file?

A prompt is a temporary instruction typed into a chat window that disappears when the session ends. An instruction file such as AGENTS.md or CLAUDE.md is a permanent document stored in your workspace. Compatible tools load it automatically at the start of every session, ensuring consistent rules without manual copying.

Do all AI tools read AGENTS.md automatically?

The filename varies by tool. Developer tools and autonomous coding agents like Cursor, Windsurf, Codex, and Claude Code look for AGENTS.md or CLAUDE.md by default. For chat workspaces like ChatGPT Team or Claude Projects, you paste the exact same 7-section markdown text into the workspace project instructions. The filename changes, but the operating manual remains identical.

How long should an AGENTS.md file be?

Keep it under 100 lines (around 500 to 800 words). If your instruction file is too long, the AI dilutes critical rules. Detailed procedures belong in linked Standard Operating Procedures (SOPs), while AGENTS.md provides the high-level behavioral contract.

How often should we update our instruction file?

Update it whenever your core offers change, when you add a new operational constraint, or when an AI agent makes a repeated mistake. Because it is a single file, modifying it once updates the behavior of all future sessions.

Want this inside your own business?

Simple AI Studio runs a hands-on implementation bootcamp for founders and small teams. You leave with a working AI system, not slides.

Keep reading