# How do you turn a vague idea into a clear question for an AI?

Date: 2022-05-06 · Author: Robert Scales · https://guidetoartificialintelligence.com/en/learn/turn-vague-idea-into-clear-question-ai/

**Answer:** Sharpen the idea in passes before you ask: say who it is for, what result you want, what limits apply, and what shape the answer should take. Each pass removes a guess the model would otherwise make, so the reply moves from generic advice to something you can use.

Most of us arrive at an AI tool with a feeling rather than a question. "I need help with marketing." "Something about my newsletter." "Ideas for the shop." Type that into the GPT-3 Playground and you will get an answer, but it will be the answer to an average version of your question, asked by an average person, in an average situation. That is rarely the help you needed.

The fix is not a clever formula. It is doing a little of the thinking before you ask, so that the model has to guess less.

## Why vague prompts get vague answers

A language model continues text in the most plausible way it can. Researchers Laria Reynolds and Kyle McDonell put it well in a paper on prompt design last year: the model's reply reflects all the ways *anyone* might plausibly continue your words, not the way *you* would. An open-ended prompt can therefore drift in several unrelated directions, as though different writers had picked it up.

Even models trained to follow instructions run into this. OpenAI's InstructGPT models, now the default in its API, were trained with people who read prompts and tried to work out what the writer wanted. The paper notes that a prompt's intention can be unclear or ambiguous, and that labelers were told to skip prompts where the task was very unclear. If trained humans struggle to read a vague request, a model will too.

## The worked example: "help with marketing" in four passes

Start with the honest first draft:

```prompt
Help me with marketing.
```

**Pass 1: Who?** Name the business and the audience. Who is selling, and to whom?

```prompt
I run a small bike repair shop in a mid-sized town. Most customers are commuters and parents with kids' bikes. Help me with marketing.
```

**Pass 2: What goal?** "Marketing" could mean a logo, a website or a sale. Pick the result that matters this month.

```prompt
I run a small bike repair shop in a mid-sized town. Most customers are commuters and parents with kids' bikes. I want more spring tune-up bookings in April and May, which are usually slow until the weather turns.
```

**Pass 3: What limits?** Constraints keep ideas realistic.

```prompt
...I have no ad budget beyond $200, I work alone, and I can spend about two hours a week on promotion. I already have an Instagram account with 400 followers and a list of 150 customer emails.
```

**Pass 4: What output?** Say what you want back and how it should look.

```prompt
...Suggest five low-cost ideas to fill tune-up bookings. For each, give one sentence on what to do, the time it takes per week, and a rough cost. Put the cheapest idea first.
```

Put together, the final prompt runs to just over 100 words. The model no longer has to invent your business, your budget or your goal; it only has to do the part you actually want help with.

| Pass | Question you answer | What the model stops guessing |
|---|---|---|
| Who | Who is this for, and who is asking? | Industry, audience, scale |
| Goal | What result do I want? | Which of many "marketing" tasks |
| Limits | What can't change? | Budget, time, tools |
| Output | What should the answer look like? | Length, format, order |

> **Try it:** Write down a vague request you've been meaning to ask. Run it once as is, then run it again after each of the four passes. Note which pass changed the answer the most.

## A useful shortcut: let the model ask you

If you are not sure what matters, ask the model to interview you first. With an instruction-following model such as `text-davinci-002`, a prompt like this is a good starting point:

```prompt
I want help planning marketing for my small business. Before giving any advice, ask me five short questions that would help you give specific suggestions. Ask them as a numbered list.
```

Answer the questions, paste your answers below them, and ask for the plan. You will often discover that you didn't know your own answer to one of the questions, which is useful to learn before spending money.

## What details does a good question include?

A good question carries the context a helpful stranger would need. In practice that means **who** it is for (you, your customers, a ten-year-old), the **goal** in one sentence, the **facts** the model cannot know (your numbers, your location, what you've already tried), the **limits** (budget, time, length, tone), and the **form** of the answer (a list, a table, a draft email).

It also helps to say what you will do with the answer. "So I can decide by Friday" or "to paste into a flyer" tells the model which details matter most.

Finally, check the question for hidden assumptions. The InstructGPT paper reports that even the improved models sometimes accept a false premise in a prompt and carry on as if it were true. If your question assumes something you haven't checked, such as "why are tune-ups my most profitable service?", the model may simply explain the claim rather than question it. Check the facts in your question first, or ask the model to point out any assumption it cannot verify.

## Sources

- Prompt Programming for Large Language Models: Beyond the Few-Shot Paradigm: https://arxiv.org/abs/2102.07350
- Training language models to follow instructions with human feedback: https://arxiv.org/abs/2203.02155
- Aligning Language Models to Follow Instructions: https://openai.com/index/instruction-following/