How to write an AI SOP.

Every small business has SOPs. Most of them are about how to train a new hire to answer the phone, pack an order, or open the store in the morning. Here is a thought that will save you hundreds of hours over the next three years: AI is a new hire. You need to train it the same way.

The difference between business owners who get great results from AI and business owners who get mediocre results is not the model they use. It is not how "smart" their prompts are. It is that one group treats prompts like instructions and the other treats them like standard operating procedures. An instruction gets used once. An SOP gets used fifty times, by multiple people, and produces a consistent output every time. That is the difference between a one-off miracle and a reliable tool.

This is the foundational piece of The Prompt Library. Everything else in this series assumes you have internalized what follows.

Treat prompts like SOPs, not like wishes

If you have ever sat at a keyboard and typed some variation of "Can you write me a follow-up email to the customer who hasn't paid their invoice?" — you have given AI an instruction. It will produce something. It might be great. It might be generic. The next time you ask, you will get a different answer, in a different tone, maybe with a different structure. Your output is a coin flip.

An SOP removes the coin flip. It says: here is the role I want you to play, here is what you need to know about my business, here is the exact task, here are the boundaries, here is the shape of the output I expect, and here are two examples of what a good version looks like. Run that every time and you get the same kind of output every time.

The six parts of a good AI SOP

Every production-quality prompt has these six sections. If one is missing, you are going to see drift in your results.

1. Role. Who is the AI being in this conversation? "You are a senior customer-service agent at a plumbing company." "You are a copywriter who specializes in short LinkedIn posts for professional services firms." A role gives the AI a voice and a perspective.
2. Context. What does the AI need to know about the situation and about your business? This is the part most people skip and it is the part that matters most. The AI is not psychic.
3. Task. What specifically are you asking it to do? Not "help me with" — do what.
4. Constraints. What are the rules? Length. Tone. Things it must include. Things it must never include. Who it should not sound like.
5. Output format. What shape should the result take? Plain email? Bulleted list? JSON? Three options to choose from? Be explicit.
6. Examples. Show, don't just tell. One or two examples of good output does more than a paragraph of description.

A working example: the payment-follow-up SOP

Here is an SOP you can steal, adapt, and start using this afternoon.

Notice what this SOP does. It specifies the role, loads the context, defines the task, sets constraints that prevent the AI from doing things off-brand, locks the output format, and tells it exactly what inputs to expect. A new employee running this prompt with your customer data pasted into the INPUT block will produce an email indistinguishable from one you wrote yourself — and will produce it in twenty seconds.

Where to store them and how to version them

Prompts are not supposed to live in your head, in a Slack thread, or in the bottom of a Google Doc labeled "random stuff." They are assets. Treat them like it.

Create a folder — call it "Prompt Library" if you want — in whatever document system your business already uses. Notion, Google Drive, Dropbox Paper, Confluence. One document per prompt. At the top of each document, write:

• Prompt name and version number (Payment Follow-Up Email · v1.2)
• What it's for (one sentence)
• Who owns it (one person on the team is accountable for keeping it current)
• Last tested (date, and how many runs)
• Change log (what changed in each version and why)

Then paste the prompt itself. When someone updates the prompt, the version number increments and a line gets added to the change log. Now you have an auditable trail of how your prompt evolved and why.

Common mistakes that kill consistency

Giving the AI no role. Without a role, the AI defaults to its most generic voice. "You are a helpful assistant" is not a role. "You are the customer-service lead at a plumbing company" is a role.

Using "help me" instead of specifying the task. "Help me write an email" is an invitation to produce mush. "Draft the email" is a directive.

No negative constraints. Telling the AI what not to do is often more valuable than telling it what to do. What phrases should it never use? What should it never assume?

Never updating. Your business changes. Your voice evolves. Your products update. A prompt that was great six months ago may now be producing outputs that reference things you no longer do. Review every prompt quarterly.

The payoff

Here is what changes when you run your business on documented AI SOPs instead of ad-hoc prompts.

Output quality becomes consistent, which means you can delegate AI work to team members without babysitting them. Onboarding time drops, because a new hire can sit down on day one and produce acceptable emails using the same SOPs you use. When a new AI model comes out, you test your SOPs against it once, adjust where needed, and ship — instead of starting over. And you build a library of prompt assets that are genuinely yours — a competitive moat that grows more valuable every time you add to it.

That is why this series exists. Every article in The Prompt Library is another SOP you can drop straight into your business. The next one to read after this, if you have never thought hard about prompt structure before, is Viki on The Anatomy of a Good Prompt.