Level 1: Prompt Engineering - Stop Asking Questions Like Google
Published April 28, 2026 · The Human Question · By Rob Gonzales, CPA, PhD
“Hey, can you go get that thing from that place?”
If you said that to a person, they would probably ask you what you meant. If you said it to AI, it would just guess. And it would guess wrong, in ways you would not catch until you needed the answer.
From buttons to language
For most of the time we have been building machines, we have made them do things by touching them. We pushed buttons. We moved a mouse. We pressed keys. Every interaction was physical, and the machine did exactly what we told it to in the most literal sense, because it had no choice. There was no interpretation.
Now we type or we speak, and the machine has to read our mind. Except it cannot. It only sees the words. The unspoken layer that humans take for granted, the tone, the inflection, the shared history, the small social signals that tell us whether “sure” means yes or means I would rather not, none of that exists for an AI. Whatever you do not say, does not happen.
That is the shift. The interface is now language. And the responsibility moved from clicking the right thing to saying the right thing.
Every prompt is an error problem
When you write a prompt, you are setting up a small prediction problem. The AI produces an output based on what you typed. The gap between what you actually needed and what it produced is the error. Prompt engineering is the discipline of reducing that error before you press send.
Watch what happens when the work is not done.
Bad prompt: “Summarize this.”
Five questions surface immediately. Summarize what, exactly? For whom? In what format? At what length? With what emphasis? The AI will guess at all five. On a good day it will guess right on two of them. The other three are now your error, and you may not notice until you are quoting it in a meeting.
The five parts of a real prompt
I recently had a few clients in the test-prep and audit-content space ask me to help their teams use AI without producing garbage. The structure I taught them, and the same one I teach my students at Fairfield, has five parts. Each one exists for a specific reason. Skip one and the error goes up.
1. Role — tell the AI who it is.
2. Objective — tell it what you are trying to accomplish.
3. Context — tell it about the world the work lives in.
4. Task — tell it exactly what to do with the context.
5. Audit — check the output before you trust it.
Role
“You are an experienced staff auditor performing preliminary analytical procedures.” This anchors tone, depth, and assumptions. Without a role, AI defaults to a generic, middle-of-the-road answer. Roles narrow the lens.
Objective
“Your goal is to identify potential audit risks suggested by the ratio trends.” AI optimizes toward goals. If the goal is fuzzy, the optimization is fuzzy. Vague objectives produce vague output, every time.
Context
“The client is a mid-size retail company with declining margins and recent ERP system changes. The financial statements and prior-year memo are attached.” Context is the world the AI is reasoning inside. Same task, different context, different answer. More errors come from a missing context than from anywhere else.
Task
“First, compute the current ratio for FY2025. Then compare it to FY2024 and identify any meaningful change. Tell me whether that change suggests liquidity risk that warrants follow-up audit procedures.” This is the literal instruction. Specific. Sequential. Bounded. Notice how different that is from “look at the ratios.” The task tells the AI exactly what to do with the context you just gave it.
Audit
“Show me the specific line items you used to compute the current ratio. Walk through the calculation step by step.” AI hallucinates. Sometimes confidently. The audit step turns the output from a black box into a workpaper you can defend. One of my best students once told me:
“I used AI to help me prepare the text for my tickmarks, but I had to do the testing myself.”
That is the move. AI drafts. You verify. The professional judgment never leaves your desk.
Before and after
Here is the same task written badly and then written well.
Before
“Help me with my audit.”
After
“You are an experienced staff auditor performing preliminary analytical procedures for a mid-size retail client with declining margins and recent ERP changes. Your goal is to flag liquidity risk that warrants follow-up procedures. I will upload the financial statements next. Wait for them, then compute the current ratio for FY2025, compare to FY2024, and identify any meaningful change. After your conclusion, walk me through every line item you used.”
The first prompt produces noise. The second one produces a workpaper.
Five rules to take with you
1. Be specific. Replace “thing,” “stuff,” and “this” with their proper names. Vague words guarantee vague answers.
2. Define the output. Bullet points or a paragraph? Two hundred words or eight hundred? A table or prose? Say it.
3. Provide context. What does the AI need to know that is not obvious from the question alone?
4. Assign a role. Even one sentence sharpens everything that follows.
5. Audit before you trust. Always ask the AI to show its work. Always check it.
The real lesson
Prompt engineering sounds technical. It is not. It is communication.
What we are really learning, when we get good at prompting, is how to be precise. How to say what we mean. How to leave fewer things unsaid, because we know the listener will not fill in the blanks.
The thing about working with AI is that it does not know what you meant. It only knows what you wrote.
Most of us were never taught how to be that clear. AI just makes it impossible to ignore.
Next up: Level 2 — Context Engineering: How AI Becomes a Co-Worker
thehumanquestion