Working with machines

How I work with AI now: talk to it, and make it draw

I talk to the machine instead of typing at it, and I make it answer on a canvas instead of in a chat. Both habits beat any model upgrade I have tried.

Vineet Sriram · draft v3 Draft · for review


Everyone asks which model I use, and the question misses where my gains actually came from, which was a pair of habits. I talk to the machine instead of typing at it, and I make it answer on a canvas instead of in a chat. The rest of this post explains why both habits work, with examples.

The fifteen-minute Jira ticket

A colleague in London needed his app, which lives in a different AWS account, connected to my service. Cross-account connectivity means PrivateLink, and PrivateLink meant his DevOps team needed the endpoint-service details from my side for both dev and prod. He had filed the ticket. I am a lead with a full plate, and this was exactly the kind of favor that eats an afternoon.

Here is everything I actually knew. Six years ago we did this for another account, so an endpoint service already exists. Somewhere there is a Jira ticket I filed at the time.

My head had kept the gist and nothing else. The service names were in an old Jira ticket, the discussion around them in Slack, the live values in two AWS consoles behind Okta, and my memory of all of it was six years stale. Every one of those screens fights for attention with things that were never my problem.

So I said it out loud, rambling and hedged, exactly as unsure as I was: "Six years back or so we created an endpoint service. There should be a Jira ticket from around the same time. Find it, use the AWS CLI to pull the endpoint service details for dev and prod, and fill in what their DevOps team asked for."

My agent is wired into Slack, Jira, and AWS. It found the six-year-old ticket, queried both environments, matched the old ticket to the live values, and put everything on one HTML page with my idea at the top and the verified details beneath it. My only remaining job was the one that should be mine: open the console and confirm the resources existed. I gave a thumbs up and it filed the ticket.

Fifteen minutes, most of it spent verifying.

Habit one: speak, don't type

When you type to an AI you close every thought before you commit it. You think a little, condense it yourself, squeeze it through the keyboard, and only then think again. The loop closes early every time because typing forces it to, and while your hands transcribe, the thinking dims. That last part is measurable. Ronald Kellogg's writing studies at Saint Louis University documented composition and transcription competing for the same working memory across the 1990s.

Speaking removes the tollbooth. You think in open loops ("maybe this, actually no, wait, before that") and nothing has to land before the next thought starts. You leave branches open. You can forget details while you wander, because the machine remembers everything and condenses it afterwards. The condensation still happens. It happens once, at the end, performed by something with perfect recall of what you said.

The numbers are lopsided. Speech runs at roughly 150 words a minute and typing at roughly 40. Caltech measured deliberate human thought at about ten bits per second, one thin serial stream out of a massively parallel brain. Typing throttles that stream even further. Voice is the widest channel you own, so it is strange to route your thinking through the narrowest one.

People expect their ramble to be garbage. That is the machine's problem now, and an AI holding the full transcript reconstructs what you meant almost unreasonably well. Kleist wrote an essay about this in 1805, arguing that the thought completes itself in the speaking. He was two centuries early to a good workflow.

Habit two: make it answer on a canvas

The second habit concerns the return path. A wall of chat text is a terrible way to see an idea, because we are visual animals reading a serial stream. So I stopped accepting text as the answer. The agent's job is to articulate what I said visually, as an HTML page or an SVG flowchart, with my idea and its details in one place I can look at.

That PrivateLink page held an interpretation, with the idea, the old ticket, the per-environment values, and nothing extraneous laid out together. When the AI must draw your rambling as a diagram, you find out instantly whether it understood you. You also find out whether you understood yourself. A wrong box is visible in a way a wrong sentence never is.

Then you iterate the way you started, by talking. "No, not that branch. Split that box into dev and prod." The diagram updates while you watch, and the iteration moves at the speed of talk on an artifact you can actually see.

A website is nothing but an interpretation of your idea. So let the machine build you one, every time.

The proof is a website that explains itself

I wanted to show this instead of claiming it, so I made the argument in the form it argues for. The result is an animated essay about talking to agents, produced by talking to agents. A particle head thinks in loops. Typing squeezes each loop through a gate into a thin dotted line, while speech overflows the silhouette toward a listener that condenses the whole wandering field and then draws it as a real SVG flowchart, live in the page.

The piece, The Shape of a Spoken Thought, running live. An act scrubber sits at the bottom, and the scientific claims are cited in the frame, including Zheng and Meister on the ten-bits-per-second bottleneck and Foster and Wilson on the brain replaying thoughts in reverse when you say "wait, before that." The animation keeps its black canvas because it is an artwork.

Making it worked the same way. I spoke the idea in loops while agents built three competing visualizations in parallel. One generated the reference silhouette for the head and another traced it into geometry. I looked at each round, said what was wrong, and the piece changed under my eyes. The whole thing took an evening, and I spent most of that evening judging results instead of producing them.

Is it 10x?

I have stopped caring about the multiplier. Here is what I can defend. Work that used to take an afternoon of context-switching now takes minutes of talking plus a verification pass. Ideas that used to die as half-typed notes become things I can look at the same day. The change I value most concerns attention. The machine does the retrieval and the articulation, and I do the judgment: is this right, and is this what I meant.

Keep the gist in your head and say it out loud. The machine will show you what you meant.


Draft for review. The research notes, the case study, and the animation source live in the thought-loops project.