The first thing I teach, and it isn't prompting

The first thing I teach is not a prompt.

That surprises people, because prompting is what they expect. They arrive wanting the sentence that makes the machine behave. They want the formula. "Act as..." "Think step by step." "Ask me questions before you answer." "Write in the style of..." They have seen lists of magic phrases and want to know which ones are real.

Some are useful. Some are harmless. Some used to work better than they do now. Some create longer answers without creating better answers.

None of them is the foundation.

The foundation is verification.

Before I care how someone asks the tool, I care whether they know what to do with the answer. Can they tell when it has invented a source? Can they spot when it answered a different question? Can they separate a polished paragraph from a correct one? Can they decide when the output is safe to send, when it needs checking, and when it should be thrown away?

If they cannot do that, a better prompt just helps them produce mistakes with better formatting.

Prompting is not useless

I use prompts all day. I write instructions carefully. I give context. I ask for structure. I tell the model what good looks like. I ask it to compare options, find gaps, and revise against constraints.

Prompting matters.

But prompt tricks are not stable enough to be the first lesson. Models change. Tool interfaces change. Memory changes. File handling changes. Reasoning modes change. What worked last month may still work, may work differently, or may be unnecessary because the model now does it by default.

Research on chain-of-thought prompting has already shown this pattern. The old advice to ask a model to think step by step can help some tasks on some models, but it can add time, increase variability, and offer little benefit on models already built to reason. Research on few-shot prompting shows a similar caution: adding more examples can improve results until it does not, and in some cases too many examples make performance worse.

The lesson is not "never prompt." The lesson is that prompting is a steering method, not a safety system.

A person who understands verification can survive a model update. A person who only memorized prompt recipes has to start over.

The answer is not the work

AI makes a dangerous kind of first draft. It looks finished.

That is what makes it useful and risky at the same time. A rough human draft usually announces itself. It has gaps, awkward transitions, missing pieces. An AI draft often arrives clean. It has headings. It has confidence. It sounds complete.

The polish can trick you into skipping the work.

In low-risk writing, that may not matter much. If you ask for a friendlier reminder email and edit it before sending, the downside is limited. If you ask for a summary of your own notes, you can compare it against the notes. If you ask for ideas, you know they are only ideas.

The danger rises when the output starts carrying facts, calculations, citations, client commitments, legal language, medical interpretation, financial assumptions, or instructions that can change a live system.

At that point, the answer is not the work. The review is the work.

The verification habit

The habit is simple to describe and harder to keep.

First, decide what kind of answer you are asking for. Is it a draft, a calculation, a recommendation, a summary, a classification, a plan, or code? Each type fails differently.

A draft can be off-tone. A calculation can be wrong. A recommendation can ignore context. A summary can omit the important exception. A classification can force messy reality into a neat category. A plan can assume people, time, or tools you do not have. Code can look right while referencing things that do not exist.

Second, decide what evidence would make the answer trustworthy. If the tool cites a source, open the source. If it names a regulation, check the regulator. If it summarizes a document, search the document. If it gives a number, trace the number. If it writes code, test it somewhere safe. If it gives business advice, compare it to the actual business, not the imaginary one in the model's answer.

Third, decide who owns the final decision. This is the part people skip. The AI does not own the email after it drafts it. It does not own the client advice after it suggests it. It does not own the spreadsheet after it writes a formula. You do.

That is not a warning against using it. It is the condition for using it well.

Why this matters for small businesses

A large company can build layers around AI use. Policies, review workflows, legal review, security tools, procurement gates, data loss prevention, audit logs.

A small business usually cannot.

That makes habits more important, not less. If there is no IT department, no legal team, and no technical reviewer down the hall, the owner and staff need a simple way to slow down at the right moment.

I teach people to tag their own AI use as they work.

Green: low risk, useful for drafting or thinking.

Yellow: check before using.

Red: do not use without approval or outside help.

Green might be rewriting a public service description. Yellow might be summarizing a client conversation. Red might be uploading medical notes, confidential contracts, financial records, employee issues, or code that changes live business data.

That system is not sophisticated. That is why it works. People can remember it under pressure.

Prompting comes second

Once the verification habit exists, prompting gets much easier.

Then I can teach people to give context, set constraints, ask for questions, request alternatives, define the audience, specify format, and ask the tool to identify assumptions. Those are useful skills. They make the work faster and cleaner.

But they sit on top of the first rule: the tool can help produce the answer, but it cannot earn trust for the answer by sounding confident.

This is also why weekly guided sessions work better than a pile of prompt templates. People need to practice on their own real work. They need to see where the tool helps, where it fails, and where their judgment enters. They need repetition, not a trick sheet.

Prompting is how you talk to the tool.

Verification is how you keep the business safe.

That is why I teach it first.

If you want to build useful AI habits on real work, start a conversation.