How to dictate an AI prompt the model can actually use

You can explain a bug to a colleague in thirty seconds, then spend five minutes trying to write the prompt. Half the useful detail disappears along the way. The result is a short request that makes the assistant guess.

Voice dictation helps you get the explanation down. The prompt still needs a shape: what is happening, what should change, what must stay true, and how to check the result. You can say all four in ordinary language.

Start with the thing you can point to

“Make checkout better” leaves the assistant to choose both the problem and the solution. Start with the behavior you actually observed.

When I choose a delivery date, go to payment, and then go back, the date is blank. The other checkout fields are still filled in.

That gives the assistant a reproduction path and a useful comparison. You have not diagnosed the cause yet. Keep it that way until the code or evidence supports a diagnosis.

If you know the relevant screen or file, include it. If you do not, ask the assistant to find it. Speaking a plausible filename is not a substitute for checking that the file exists.

Add the change and the boundary

Follow the observed behavior with the result you want. Then say the exception that would make an otherwise reasonable implementation wrong.

Keep the delivery date when the customer goes back from payment. Clear it if they change the cart, because the available dates may be different. Use the existing checkout state.

The exception matters. Without it, preserving the date everywhere could introduce a second bug. This is the kind of detail people often skip when they are trying to make a typed prompt shorter.

OpenAI’s prompting guidance recommends clear, specific instructions, sufficient context, and refinement based on the response. Dictation changes how you supply those instructions. It does not remove the need for them.

Say how you will know it worked

Close with something the assistant can inspect or test.

Check the back-navigation path and the cart-change path. Before editing, tell me where the date is currently stored and which tests cover it.

Now the request has an observable outcome. The assistant has a reason to inspect the current behavior before inventing a new state system.

Here is the whole example as one spoken brief:

The checkout page loses the selected delivery date when I go back from payment. The other fields stay filled in. Keep the date on back navigation, but clear it when the cart changes. Use the existing checkout state. First find where the date is stored and tell me how you would test both paths.

This is a fictional teaching example, not a claim about a customer project or a measured improvement in model performance.

Review the details that speech can get wrong

Before sending, read names, numbers, paths, and negations. “Do not change the public API” is a very different instruction if “not” disappears. A misheard file path can waste the next few minutes even when the rest of the prompt is excellent.

Paste error messages and exact identifiers when that is easier. You do not have to dictate everything to make dictation useful. Attach the screenshot or file the assistant needs; talking about a document does not give it access to that document.

In Vokalu, use ordinary dictation to enter the prompt into the receiving app. Command Mode is a separate action for rewriting selected text. Review the result before submitting, especially if you used a more aggressive formatting level.

Keep the next prompt narrow

When the assistant asks a useful question, answer it. When it makes a wrong assumption, correct that assumption directly. Repeating the entire brief can make it harder to tell which part changed.

Try this on one real task: dictate the first brief, inspect the assistant’s understanding, and note what you had to clarify. The useful measure is whether the assistant had enough information to do the work. The word count of the prompt is secondary.

For the Windows shortcut and a worked example, see voice dictation for AI prompts.