Try Asking AI Out Loud

Imagine opening the fridge, looking at what's there, and still having no idea what to cook. You tap the microphone in your AI chat and say:

"I've got pasta, tomatoes and cheese. It's for two people. I'm tired, so something easy. Actually, I don't want to go shopping. What can I make with what I've got?"

You've explained more than the contents of your fridge. Dinner needs to feed two people, take little effort, and use ingredients you already have. Those details change what a useful answer looks like.

This is why I think speaking to AI is worth trying. When we type, we sometimes shorten a whole situation into a neat request: "Give me a dinner recipe." The things we leave out can be exactly what would help the AI suggest something we can use.

Speaking gives those details room to come out. You can pause, change your mind, or add something you nearly forgot. You don't need perfect grammar or technical language. Explain it as you would to someone sitting beside you. Include what matters; you don't need to keep talking just to make the request longer.

Suppose the AI suggests a meal, but the recipe needs several pots and pans. Looking at it, you realise that cooking is only half the job. You tap the microphone again:

"That sounds nice, but there's too much washing up. Can you suggest something I can make in one pan?"

The answer helped you notice another preference. You added it, and the conversation continued. You didn't have to work out every requirement before you began.

An AI assistant can often make a start with an incomplete explanation. That's useful, though it can also fill the gaps with assumptions and take the job further than you intended. You ask about dinner tonight and get a week's meal plan, complete with a shopping list.

It can feel like asking someone to paint one wall and finding they've painted the whole house.

You can bring it back with ordinary words:

"Only dinner tonight. One simple idea, using what I already have."

You can speak casually and still be particular about what you want. Read the answer, check whether it helps, and explain what needs to change. The same back-and-forth can help with research or a longer project, while you stay involved as the work develops.

That back-and-forth works because you're there to notice what's missing and explain what needs to change. But imagine building a dinner-planning service that answers thousands of requests, or eventually a million. You won't be there to correct each suggestion.

If you're building a service that uses AI to handle requests automatically, the instructions need to carry more of the work. Our dinner service needs to know how to use someone's ingredients and serving count, when to ask a question, and what to do when information is missing.

Designing those instructions, testing them with different requests, and improving them when they fail is part of prompt engineering. That work becomes especially important when a product keeps doing a job without you checking every attempt.

Testing our dinner service might reveal that its easy meals still leave people with too much washing up. You could revise the instructions and try again. Even carefully tested instructions cannot guarantee that every answer will be right.

You can give careful instructions by speaking, too. The microphone changes how you enter your words; prompt engineering is about how you design and test the instructions.

When you're staying involved, you can start with an ordinary explanation and improve it together.

Try that with something you need help with in ChatGPT or Gemini. Press the microphone and explain what you're trying to do. Read the answer, then say what needs to change.

You don't have to find the perfect words before you begin.