Most bad AI answers start with missing context.
You know what you want.
You know the examples. You know what failed. You know the small detail that changes the whole task.
Then you cut all of it down to three neat sentences.
The AI gives you a generic answer.
So stop trying to write the perfect prompt. Talk through the full problem instead.
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…
— Andrej Karpathy (@karpathy) July 21, 2026
Messy speech has more signal
A long voice note keeps the parts a polished prompt removes.
It keeps the example that shows what “good” means.
It keeps the exception that changes the rule.
It keeps the moment you say, “Wait. That is not quite right.”
Those corrections matter. They show the AI where your real goal sits.
What sounds messy to a person can look like useful evidence to a model.
Vibe coding starts with a rant
Karpathy gave us “vibe coding.”
But the code comes after you explain the product.
You still need to say what it should do. You need to share why it matters. You need to name the trade-offs.
Voice makes that easy.
OpenClaw creator Peter Steinberger uses long spoken prompts to build software. In his talk with Lex Fridman, he jokes that his hands are too precious for writing. He once used voice so much that he lost his voice.
That story changed how I worked.
I had built a small dictation tool for myself. Soon I used it 20 times a day.
I could explain a bug. Write a Slack message. Plan a feature. Or turn a rough idea into a prompt for Codex.
Think about Iron Man.
Tony Stark does not type a perfect prompt into Jarvis. He talks. Jarvis asks questions. They build the answer together.
That is the interface AI has been moving toward.
This is truly my top point of advice for people trying to level up their usage of LLMs. They are *very* good at understanding your semi-coherent thoughts. Particularly when you ask them to *ask you questions* about those thoughts.
— Matt Stockton (@mstockton) July 21, 2026
Use this 4-step loop
A rant works best when the AI talks back.
- 01Dump
Speak for five or ten minutes. Include the background, examples, objections, and the parts you have not figured out.
- 02Reflect
Ask the model to state what it thinks you mean, what appears important, and where your thinking conflicts.
- 03Interrogate
Tell it to ask questions. A good follow-up often extracts the one constraint your first rant only implied.
- 04Aim
Choose the finished work: a product brief, email, plan, post, script, task list, or prompt for the next agent.
Voice gives AI 3× more to work with
A Stanford study found that English speech input was about three times faster than smartphone typing.
That does not mean every person speaks three times faster in every app.
It proves the basic point: speech can move a lot more information.
And more information means fewer guesses.
When typing feels slow, we cut the history. We cut the odd detail. We cut the “this sounds strange, but…” part.
That is often the part the AI needs most.
Better context beats a better prompt.
Save the rant
Not every rant needs to become a blog post.
Some rants help you think.
Some become emails. Some become product plans. Some sit for a month before they become useful.
So keep the source.
Today it becomes a Slack message. Next month it may hold the missing detail for a product choice.
You should not need to remember which AI chat heard it first.
When the thought feels larger than the prompt, say the whole thing.