Voice mode

14 October 20263 minute read

One pattern I find useful for working with LLMs is a nice long ramble session. Sometimes the LLM needs more bits to understand what you're trying to achieve, but you're too lazy to type them. In these cases I like to lean back, switch to /voice and just ramble for like 10 minutes, total mess, anything goes, full stream of consciousness. Sometimes I declare it up top, something like "switching to speech recognition sorry for any typos...". Sometimes I turn it into a small interview of a few turns. But I find that the LLMs are somehow very good at reconstructing long incoherent rambles and often their echo of your own tangle of thoughts comes out quite a bit cleaner than what you started with. The result is that you improve the mind meld and have to correct things less from that point on.

Andrej Karpathy (@karpathy), 22 July 2026. Source: supplied post screenshot.

View supplied post

A man in a red shirt plays a small guitar in bed beside a lit lamp and a city window. — Her (2013). Supplied image from Becoming Engineers, slide 33.

In this guide

Choose the interaction

Live conversation is an exchange. You speak, the model responds, and you can ask a follow-up question. Use it to examine an idea, explain a problem or rehearse an argument. State what you want from the exchange before you begin.

Dictation turns speech into written text. You can dictate into a document or compose a message for a model. You then decide when to review, edit or submit the text. This separates getting the words down from asking for a response.

Check which interaction your chosen tool provides. A microphone button can mean either one. If you need a transcript, check that you can read and retain it before starting.

Think aloud

State the question you want to resolve. Then talk through the background, your current view and what remains uncertain. Let the first pass include unfinished ideas. Ask the model to reflect back what it understood and identify the gaps.

Ask for a specific response. A useful instruction might be:

“I am weighing two approaches. Ask me one question at a time about the assumptions behind each. After we discuss them, summarise the trade-off.”

Keep your observations separate from your guesses. If the model introduces a fact, mark it for verification. At the end, ask for the unresolved questions and your next action in writing.

Draft by speaking

Name the reader and the purpose of the draft. Speak through the main point, the evidence and the action you want the reader to take. A rough first pass can contain repetition and unfinished sentences; edit these in the next step.

For dictation, begin with one paragraph. Read it back before adding more. Correct any names or technical terms that the transcript has captured incorrectly.

If you ask a model to edit the transcript, set a clear boundary:

“Turn this transcript into a short draft for a colleague. Keep my meaning and the facts I supplied. Flag gaps and ambiguous passages for me to resolve.”

Read the result against the transcript. Restore any qualification that changed the meaning. Add evidence where it is needed and remove claims you cannot support.

Review the transcript

Speech can be clear in the moment and incomplete on the page. Review the written record before you use it as a draft, instruction or decision note.

Check recognition. Correct names, numbers, dates and specialist terms against your source.
Check meaning. Resolve unclear references, missing conditions and sentences with more than one possible reading.
Check attribution. Separate your statements from the model’s suggestions and verify any new factual claims.
Check the action. State what happens next, who owns it and which questions remain open.

Choose a setting where you can speak without exposing private information or recording other people. Check the tool’s recording and data settings for the material you intend to discuss.

Finish with a written record you can inspect. Keep the useful reasoning, confirm the facts and decide the next step yourself.