BackAI Prompting: It’s You, Not the Tool
When I first started using AI, I often felt like I’d get responses back that weren’t satisfying to me. I’d blame the tool, thinking it wasn’t advanced enough to understand me or had hallucinated information that I hadn’t told it. It wasn’t until I learned how to write a good prompt that I realized how much of the onus was on myself to prompt the AI with all of the potential information, instructions, and guidelines as possible in order to get back exactly what I was looking for.
We recently helped run a hands-on AI workshop for managers at a contract manufacturer in a regulated industry. As a product manager, much of my job is turning vague requests into clear requirements. When I interacted with a room full of people who were new to using AI, I realized that they needed to learn that same skill. The room was full of experts in their own work, who are used to interacting with their colleagues, who are also experts in the field. They needed to learn how to interact with an AI who didn’t have that context, even if its responses made it seem like it did.
The Workshop
The session was built around the attendees' real work. Managers set up shared workspaces around their own department workflows, turned repeatable tasks into reusable instructions, and mapped where AI could help and where it couldn't. Before any of that, though, we spent time on the thing that most determines whether an AI assistant is useful: the prompt. That segment covered three ideas: the five parts of a good prompt, letting the AI help you write the prompt, and a short review loop for staying in control of what comes back.
The Workout Plan That Went Too Far
A clear example of why this matters came from participation of an attendee in the room. During one of the hands-on exercises, an attendee tried a personal request to see how the AI would handle it:
“Make me a workout plan. I want to have a variety of strength training, cardio, and pilates in my plan.”
It's a perfectly reasonable request, and it's how most of us would ask a friend who happens to be a trainer. What came back was a seven-page document: five days a week of exercises, laid out over a twelve-month timeline, with different levels of strength training based on a potential user's weight.
None of that was in the prompt. To produce that plan, the AI had to decide on its own:
- How often the person wanted to work out (five days a week)
- How long the plan should run (a full year)
- How much detail they wanted (seven pages)
- What mattered about the person when setting intensity (their weight)
It also had to work around everything it was never told: what the person's goal was, how fit they were starting out, how much time they had per session, and what equipment they could use.
The AI didn't do a bad job. It did exactly what it was asked, which was to fill in every blank by itself. The problem is that a polished, seven-page plan looks authoritative. It's easy to miss that most of it rests on guesses.
The same thing happens at work, with higher stakes. The example on our slides was a one-line request for a night-shift handover. Given no other information, the AI produced a generic template and, where it filled in specifics, invented plausible-looking owners and times. When it was done, there was no easy way to tell which parts came from the actual shift notes and which were made up. In a workout plan, a wrong guess wastes your time. In a shift handover, it can get someone in trouble.
The Five Parts of a Good Prompt
The fix is to leave fewer blanks, so that the AI doesn’t have to make any assumptions. A good prompt answers five questions the AI would otherwise guess at. Here's how they reshape the night-shift handover: same task, same request at the core, with five short lines added.
1. Goal: what are we doing? "You are helping me write the shift handover for the QC lab night shift." One sentence already narrows things a lot. Not a handover for any team anywhere, but for one lab and one shift. That alone gets you out of generic-template territory.
2. Context: who is it for? "The reader is the incoming supervisor. They have five minutes before the shift starts." This is the line people skip most often, and it's the one that fixes length. Once the AI knows the reader has five minutes, it writes for five minutes.
3. Source: where should the facts come from? "Use only the notes below. If something is not in the notes, write 'Not in notes.'" For work like this, it's the most important line in the prompt. It's what stops the invented owners and times. You're telling the AI where the facts come from, and giving it permission to leave a gap instead of filling it.
4. Format: what shape should it take? "Three sections: Done, Pending, Questions. Under 200 words. Plain language." You already know what a good handover looks like. Describe the shape, and you won't have to restructure the draft afterward.
5. Checks: what should be double-checked? "Before you finish, list any owner or date you were not sure about." This turns the AI into a reviewer of its own work. Instead of rereading the whole draft with suspicion, you get a short list of exactly where to look.
Then you paste your notes underneath and send it. If you need a memory hook, the five parts follow the questions you'd answer for a new colleague: what are we doing, for whom, from what, in what shape, and what should you double-check? Writing them takes about two minutes. That saves the twenty you'd otherwise spend fixing a draft that guessed wrong.
Let the AI Interview You
Five parts is still a lot to remember on a busy day. The shortcut most people don't know about is that you don't have to write the prompt yourself. You can ask the AI to interview you first:
I need to do [task] for [reader]. Before you write anything, ask me up to three questions, all at once. Then write the prompt I should reuse, with goal, context, source, format and checks.
"Before you write anything" stops it from jumping straight to a draft. "All at once" means you answer every question in a single reply instead of a long back-and-forth.
This is exactly what the workout request was missing. An AI that asks first would have wanted to know how many days a week you can train, how long each session is, what you're working toward, and what equipment you have. Those questions surface the details you didn't think to mention because they were obvious to you. Every one of them is a blank that would otherwise have been filled with a guess.
From there, the process is short. Test it: paste the prompt it wrote into a fresh chat and run it on real input, so you know the prompt works on its own. Fix the one thing it gets wrong; usually it's just one. Save it: keep the prompt somewhere you'll reuse it, such as a shared workspace's instructions or a saved skill, so you never write it twice.
The line I most wanted people to leave with: you don't need to be good at prompts. You need to know your own work. The AI can handle the structure. It can't know your goals, your schedule, or how your team runs, and you already do.
Four Follow-Ups That Do Most of Your Review
Once you have a first draft, most people either use it as-is or start rewriting by hand. There's a better middle step: four short follow-ups, each a sentence you can type in a few seconds.
Shape it. "Shorter, for my director. Keep the open questions." When a draft is right but pitched at the wrong reader, don't start over. Say what to change and, just as important, what to keep. Had the workout attendee typed "Cut this to one page and four weeks," they'd have had something usable in seconds.
Surface assumptions. "What did you assume that isn't in my request?" This one is simple and surprisingly powerful. Asked of the workout plan, it would have brought the five days, the twelve months, and the weight-based intensity straight to the surface.
Ask for proof. "Quote the line in the notes that supports each date and owner." This ties every specific fact back to your source. If the AI can't point to a line, treat that fact as unverified.
Find the gaps. "What's missing that I should ask before I send this?" These are the questions you'd otherwise hear from your reader after it's too late.
If you only use one, make it "ask for proof." Then do the one thing the AI can't do for you: open your source and find that line yourself. In the workshop, we closed this exercise with a single debrief question: did anyone find a claim that didn't match their notes? If you did, the process worked. Catching it before the draft leaves your desk is the whole point.
The Bigger Lesson
The workout plan is a useful example because it's so familiar. We've all asked for something in one line and gotten back an answer to a question we didn't ask. With people, a colleague usually fills the gaps with shared context. An AI fills them with guesses, and then presents those guesses with the same confidence as everything else.
Prompting well isn't a technical skill. It's a habit: give the AI the five things it needs, let it ask you questions when you're not sure what those are, and review the draft before you trust it. That habit matters for anyone using these tools, and it matters most for teams whose work ends up in front of a client, an auditor, or a regulator. The AI can handle the structure. Knowing what you actually need is still your job, and it's the part you're already good at.
