We built the asking-back part instead of teaching people to prompt

"You know the thing — for Instagram, something on-trend but not too light, some copy for the new product." Say that to a colleague and it usually lands. Paste it into an AI and what comes back is flat.

CHALTTEOK is a prompt generator built to close that gap. This is a record of how we framed the problem, and why we did not go the route of teaching people to prompt better.

1. The gap was missing conditions, not missing skill

When someone says AI is not giving them what they want, the usual suspects are the model or the question. Watching people work, it was neither. Their requests were clear enough to work between two humans.

What was missing were the conditions. Who reads this, how long it can be, what shape it has to come out in. A person fills those in, or asks. A model does not ask — it quietly fills them with its own average. That is why the output is safe and forgettable.

The problem was not a badly written question. It was that the details deciding the answer were never in the question.

2. We decided not to make people study

The common fix is to teach prompting. There is no shortage of courses and templates. But that fix comes with conditions: you have to find time to learn it, and you still have to write it all out every time. The work stays on the person.

Our product principle points the other way. When technology disappears, people can concentrate on what they are actually doing. If a tool has to be studied before it is useful, the tool is unfinished, not the user.

Instead of teaching people how to prompt an AI, let the AI ask for the conditions it is missing.

3. Ask back, without turning it into an interrogation

"Just ask them" is the easy conclusion, and building it literally turns the product into a form. Line up ten questions and people quit around the second one. So the asking has three constraints.

• Only ask what changes the outcome — if either answer produces the same result, the question is cut.
• Every question ships with a suggested answer — you pick, instead of staring at an empty field.
• You can skip all of it — anything guessed shows up in the prompt marked as an assumption.

That last one carries the most weight. Force an answer and people leave; guess silently and they have no idea why the result drifted. Leaving the guesses visible avoids both, because now only the wrong ones need fixing.

4. The output is a prompt, filled in

The conditions gathered that way come back as a prompt a person can read. The offhand request above turns into this.

Role: social copywriter for a 20s-30s audience
Tone: on-trend, still credible
Goal: 3 options of Instagram copy for the new product
Limits: two lines each, hashtags included

Showing this rather than hiding it is deliberate. When you can see which line moved the result, you can fix it yourself, and after a few rounds you start writing that way from the beginning. We decided not to teach. We did not stop people from learning.

5. No copy-paste, and no rewriting either

Making people copy the prompt into another window is one more place to lose them, so it runs where it was built. You can still send it over to ChatGPT or Claude, but the default path never leaves the screen.

The same holds when the result misses. Normally you have to describe what was wrong, in writing — which is exactly the problem we started with. Instead, you pick a reason.

• Too long · Not enough substance
• Too generic · Wrong format
• Missed something important · Not sure it's accurate

Pick one and the prompt is reworked in that direction. Don't like it, roll it back. The undo is what makes people willing to press the button at all, and pressing it is where the improvement actually happens.

6. What is still open

The hardest part is still deciding what to ask. Ask too little and the output is bland; ask too much and people give up. That line does not resolve into a rule, so we keep moving it based on how the product is actually used.

It was also never meant to be the only one. Repeated explanations, complicated setup, interfaces you have to study first — there are plenty of barriers left between people and what they are trying to do. Removing them one at a time is the plan, and this was the first item on the list.