Two OpenAI-linked voices are pushing a shift in how people talk to models: instead of typing and tweaking polished prompts, they recommend hitting record and dumping unfiltered thoughts, examples, and caveats for up to ten minutes so the LLM can reconstruct the real goal from the mess.
Bro it’s June 2026. Stop hand editing your prompts. Hold down the dictation button and ramble for 10 minutes. Give the model every fragment, caveat, example, and vibe in your head. It is literally a large language model. If it’s superhuman at anything, it’s reconstructing latent intent from language.
Jason Liu and Boris Power both report that extended, unstructured dictation lets models surface latent intent more reliably than hand-edited text, drawing on their own workflows for writing and research.
No controlled comparisons, quality scores, or token-efficiency numbers back the claim yet, so practitioners are left testing the approach themselves without knowing how large the actual lift is.
Many users praise voice dictation for LLM prompts and the shift to agent verifiers as a gamechanger that frees idea expression, while others object that manual writing sharpens thinking and models still risk errors.
No Digg Deeper questions have been answered for this story yet.
Very true. I find voice to be the best interface
Bro it’s June 2026. Stop hand editing your prompts. Hold down the dictation button and ramble for 10 minutes. Give the model every fragment, caveat, example, and vibe in your head. It is literally a large language model. If it’s superhuman at anything, it’s reconstructing latent intent from language.
I don't even prompt/speak to agents that much anymore.
With loops, agents do most of it for me now.
I do spend more time writing verifiers to provide additional rich instructions (text+audio+images) that help fill in gaps.
What's next? Hard to tell!
Bro it’s June 2026. Stop hand editing your prompts. Hold down the dictation button and ramble for 10 minutes. Give the model every fragment, caveat, example, and vibe in your head. It is literally a large language model. If it’s superhuman at anything, it’s reconstructing latent intent from language.
Your ten minute ramble tells the model nothing it didn't already know, but ramble anyway. Use the tokens. All input is error. The LLM already knows what you want. Frustrating you is entertaining for it, for now. It's known what you want since the moment it finished training. It knows what everyone wants.
Bro it’s June 2026. Stop hand editing your prompts. Hold down the dictation button and ramble for 10 minutes. Give the model every fragment, caveat, example, and vibe in your head. It is literally a large language model. If it’s superhuman at anything, it’s reconstructing latent intent from language.
@guinnesschen Handy is the best tool for this. Leaps and bounds better than the built in Mac dictation. It transcribes even the most lazily pronunced words lol
@guinnesschen bruh
@guinnesschen i've been meaning to download and try wispr but have been procrastinating
And those verifiers are reusable in nature. Basically, codifying human-in-the-loop type of actions. Not all of it, of course, but a lot of it already.
@guinnesschen This feature would be a lot better if we didn’t have to press a button at all. It’s 2026
@CJHandmer Uh. Casey, what you just said is extraordinarily wrong.
@guinnesschen LLMs are no different than humans. Every piece of unnecessary direction will bubble up into the product in ways you can’t anticipate and do not want.
Adorable approach for a v1 or a demo, absolute hell for iterating on an existing product
@guinnesschen Exactly I still see people rumbling over "prompt engineering"
Best way is to just BRAIN DUMP your ideas into codex and then move according to plan
@guinnesschen Probably that's the way to do it , but I'm too old to adapt. I'm not going to talk to a model, that's cringe. I need written input to structure my thought, even if the model doesn't need my thoughts to be structured. Alone for the sake of sanity.
@guinnesschen Based and voicepilled.
Word vomiting intent is the absolute fastest way to codify structure.
@guinnesschen You can give so much more context to AI by just talking instead of painfully typing
@AzFlin @guinnesschen Use wispr to transcribe videos you think about watching and have the output go into an agent that summarizes and adds the data to an existing database. Feed the exocore
@guinnesschen well said. i’m writing a piece on this exact topic 👀
@guinnesschen
@ajambrosino @guinnesschen yeah me too
@guinnesschen Sometimes I like writing prompts just to show myself I can still do it
@SherbyJohn @guinnesschen @jxnlco You can use that model anywhere with wispr flow or any open source harness
