How to write instructions an AI can actually follow: prompting for business owners, minus the mysticism

Most AI outputs fail because the instruction was written for a human, not a machine. Four moves — role, context, output format, examples — fix most of it.

The single most common reason AI gives a useless answer is that the instruction given to it was written for a human colleague, not a machine. A human fills in what you left out; an AI produces exactly what you described, no more. Effective AI instructions follow four moves: give the AI a role so it knows what perspective to adopt, state the context so it knows why the task matters, describe the output you want precisely — format, length, audience — and include one or two examples when format matters. That is the full framework. No jargon required.

Why the first draft usually fails

Most people start with something like "write me an email about our new service." The AI produces a generic, professional-sounding email — technically correct and completely wrong for the situation, because it had no information about your industry, your reader, your tone, or what action you wanted the reader to take. Anthropic, the company that makes Claude, describes this in its documentation as the "brilliant but new employee" problem: "Think of Claude as a brilliant but new employee who lacks context on your norms and workflows. The more precisely you explain what you want, the better the result." The fix is not a better AI tool — it is a better instruction.

Move 1: give it a role

One sentence assigning a role changes the character of the output. Anthropic's documentation states that even a single sentence makes a difference. "You are a customer-service assistant for a Thunder Bay heating and cooling company" focuses the model's tone, vocabulary, and assumptions differently than no role at all. The role does not need to be elaborate — it needs to be accurate. A practical test: if the role sentence would make sense as the first line on a job posting, it will work in a prompt.

Move 2: add context and the reason

Explaining why the task matters changes what an accurate response looks like. Anthropic's documentation states: "Providing context or motivation behind your instructions, such as explaining to Claude why such behavior is important, can help Claude better understand your goals and deliver more targeted responses." An example from that documentation: instead of "NEVER use ellipses," the instruction "Your response will be read aloud by a text-to-speech engine, so never use ellipses since the text-to-speech engine will not know how to pronounce them" gives the model enough information to generalize correctly. In a business context: "This email will go to a customer who has already complained once about delivery times" produces a different — and better — output than "write a customer email."

Move 3: describe the output exactly

Specify format, length, and audience. Anthropic's documentation converges on a consistent point: describe what to do, not what not to do. Their guidance: "Tell Claude what to do instead of what not to do." Instead of "don't be too long," write "respond in three bullet points, each under twenty words." The negative instruction leaves interpretation open; the positive instruction does not. For recurring tasks — weekly reports, quote summaries, review responses — this specificity is what makes the output consistent enough to actually use. OpenAI's prompt engineering guide makes the same recommendation, framing it as one of the core components of an effective prompt.

Move 4: use examples when format matters

Anthropic's documentation describes examples as "one of the most reliable ways to steer Claude's output format, tone, and structure." If you have a format you already like — a specific customer-email style, a report structure, a way of writing estimates — paste one example in and say "match this format." Two or three examples produce noticeably more consistent results than a written description of the format you want. OpenAI refers to this technique as few-shot prompting in its official documentation. The investment in writing two good example outputs pays off on every repeat run of a routine task.

A reusable template for any repeating business task

The four moves fit into a pattern you can reuse. Fill in the brackets for any task, then save the instruction for next time. A Thunder Bay bookkeeper drafting client follow-ups might write: "You are a professional bookkeeper sending follow-up messages to small-business clients in Northwestern Ontario. This message will go to a client whose invoice is 30 days overdue. Write a polite one-paragraph email asking for payment without making the client feel accused of wrongdoing. Keep it under 80 words. Here is an example of the output I want: [paste example]." The specifics change per task; the structure stays the same.

• Role: You are [what the AI should act as] for [your business type].
• Context: This [output] will go to [audience], who [relevant detail about their situation].
• Task: [Specific action verb] about [subject].
• Format: Respond with [format], in [length], in [tone].
• Example: Here is an example of the output I want: [paste example].

Frequently asked questions

• Does prompting work the same across different AI tools — ChatGPT, Claude, Copilot? The four principles apply to any large language model. The surface details vary: ChatGPT and Claude use different interfaces; Microsoft 365 Copilot is embedded in Word, Outlook, and Teams. The underlying approach is the same: the more precisely you describe what you want, the closer the output will be. Expect some adjustment for each tool, but the framework transfers.
• How long should a good prompt be? As long as it needs to cover the four moves — no longer. A two-sentence instruction for a simple task is fine. A 200-word instruction for a complex, recurring task is also fine. Anthropic's "golden rule": show your prompt to a colleague with minimal context and ask them to follow it. If they would be confused, the AI will be too.
• Should I save my prompts somewhere? Yes, for any task you repeat. A shared folder of tested instructions organized by task type is a practical starting point. Some AI tools offer "custom instructions" or "system prompts" that let you set a default context for every session — check whether the tool you use supports this, as it saves retyping the role and context each time.

Sources: Anthropic — Prompting best practices ("brilliant but new employee" framing, golden rule, context/motivation guidance, "tell Claude what to do" guidance, examples as "one of the most reliable ways"): docs.anthropic.com/en/docs/build-with-claude/prompt-engineering/overview | OpenAI — Prompt engineering guide (four-part prompt structure, few-shot learning): platform.openai.com/docs/guides/prompt-engineering