How to become a 10x ramble-coder
Substack ● Covered by 2 sources
A startup CTO says he stopped typing prompts to AI coding agents and now just rambles out loud instead. His argument: unfiltered talk carries way more useful context than a tidy written prompt ever could.
Based on reporting by Substack — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Andrej Karpathy recently called it his new favorite move: a long ramble session with an LLM instead of a carefully typed prompt. Moe Khalil, CTO at Webhound, has apparently taken that idea and run with it for months, to the point where he says he barely touches his keyboard anymore when working with coding agents like Codex.
His case is pretty simple. Writing a prompt forces you to compress a messy, contradictory train of thought into a few clean sentences, and that compression throws away exactly the information an agent needs. Khalil's login-screen example makes the point well: his internal monologue bounces between social login options, image choices, modal versus split-screen layouts, and a running argument with himself about whether any of it even matters. A typed prompt reduces all that to "clean, split-screen login with social auth." Everything about why he almost chose a modal, or why he cares about avoiding page redirects, just vanishes.
The more interesting tactic is what he calls multiplayer rambling. He and his cofounder record voice-memo conversations, sometimes running for hours, deliberately disagreeing with each other about product decisions, technical tradeoffs, even copywriting. They run the audio through Whisper with speaker identification and dump the transcript into Codex as ground truth. The agent isn't asked to resolve the argument, just to absorb the full context of it, and Khalil claims this makes it behave like a third cofounder who somehow already knows the team's priorities.
He also pushes back against writing feedback as tidy bullet points, arguing that voicing frustration, uncertainty, and emotion tells the agent what actually matters to you. A flat note like "the copy is too long" carries far less signal than rambling about why the copy feels sales-y and disingenuous. The setup requires almost nothing new: a dictation app like Wispr Flow or Willow, or a phone recording plus any speech-to-text model with speaker labels for group sessions. Khalil's pitch is that the friction of talking to your neighbors' annoyance included is worth it because the output finally matches what you actually pictured.
My take — AI-written commentary, not fact-checked reporting
This is basically prompt engineering rebranded as therapy, and I don't hate it. Voice input genuinely does carry more of the reasoning trail than text, so using it to feed agents context makes sense, but let's not pretend recording hours of "constructive arguments" scales to anyone without a cofounder, a Substack, and a high tolerance for annoyed neighbors.
Read more about this at: Substack