Stop Prompting. Start Onboarding: How to Train an LLM Like a New Hire

Most people use an AI model the way they use a search bar: type a few words, hope for the best, and blame the tool when the answer misses. Then they conclude the technology is overrated.

Here is a better mental model. Treat the LLM like a smart, eager new hire on their first day. They are capable, fast, and completely unfamiliar with your business. If you handed a new employee a three-word instruction and walked away, you would not expect great work. The model is no different.

Wharton professor Ethan Mollick, author of Co-Intelligence, gives the same advice: treat AI like a talented intern — delegate real work, but supervise, give context, and check the output. The teams that get value from AI are the ones that manage it, not the ones that poke at it.

1. Give it a job description, not a task

A new hire performs better when they know their role. So does a model. Before the task, tell it who it is and who it is working for.

Weak: “Write an email about the outage.” Better: “You are an IT service manager at a Denver managed services firm. Write a calm, plain-spoken update to non-technical business owners about a resolved email outage. No jargon, no blame, under 150 words.”

Role, audience, tone, and constraints. That is a job description — and it consistently outperforms bare instructions.

2. Onboard it with context

You would never ask a first-week employee to write client-facing work without background. Models are the same: they only know what you put in front of them.

• Paste the relevant background — the policy, the thread, the meeting notes

• State what the output is for and who will read it

• Include the constraints that matter: length, format, deadline, things to avoid

• Tell it what good looks like — “clear enough for a non-technical owner to act on”

OpenAI’s own prompting guidance puts this first: the single most common cause of bad output is missing context, not a weak model.

3. Show, don’t just tell

Training a new employee usually means showing them a finished example and saying “make it like this.” The technique has a formal name in prompting — few-shot examples — and it is one of the most reliable ways to improve results.

Paste one or two examples of past work you were happy with and say: match this structure and tone. The model will imitate the pattern far more accurately than it will interpret adjectives like “professional” or “friendly.”

4. Let it think before it answers

Good managers do not demand instant answers to complex questions. For multi-step work — analysis, plans, anything with reasoning — ask the model to work through the problem step by step before giving its conclusion, or to outline first and draft second.

Research on chain-of-thought prompting from Google found that asking models to reason through intermediate steps significantly improves accuracy on complex tasks. It is the prompting equivalent of “walk me through your thinking.”

5. Give feedback like a manager, not a customer

When a new hire’s first draft misses, you do not walk away and hire someone else. You say what was wrong and ask for a revision. Do the same here.

• “Good structure, but too formal — rewrite for a client we know well”

• “Cut the second paragraph and make the action items a numbered list”

• “This claim is wrong — here is the correct figure, revise around it”

Iterating in the same conversation is almost always faster than starting over. Each correction is training data for the rest of the session.

6. Write down what works

When an employee finds a process that works, you document it. Do the same with prompts. When a prompt reliably produces the report, summary, or email you want, save it in a shared library with a name and a note about when to use it.

That is the difference between one person being good at AI and the whole business getting faster. It also gives leadership visibility into how AI is actually being used — which matters for security and data handling.

One warning before you delegate

Like any new employee, the model will occasionally state something wrong with complete confidence. Verify facts, figures, and anything client-facing before it ships. And never paste sensitive data — client records, credentials, regulated information — into public AI tools your business has not vetted. Ownership of the output still sits with you.

How Entice Technology helps

Entice helps Colorado businesses adopt AI deliberately: setting up secure, business-approved tools, defining what data can and cannot be shared, building prompt libraries around real workflows, and training teams so the productivity gains do not come with a security bill attached.

Sources

Ethan Mollick, Co-Intelligence: Living and Working with AI — the “treat AI like an intern” framing: https://www.moreusefulthings.com

OpenAI prompt engineering guide: https://platform.openai.com/docs/guides/prompt-engineering

Anthropic prompt engineering documentation: https://docs.anthropic.com/en/docs/build-with-claude/prompt-engineering/overview

Wei et al., Chain-of-Thought Prompting Elicits Reasoning in Large Language Models (Google Research): https://arxiv.org/abs/2201.11903

Ready to put AI to work properly?

Entice Technology can help your team adopt AI tools securely and train staff to get consistent, reliable results — without exposing business data. Contact us to talk it through.