Conversations with Claude

Voice as an epistemic tool for thinking

Conversing with the LLM using voice. (Image courtesy of Google Gemini 3.1 Image Flash.)

We are in something of a Wild West when it comes to LLM use. No one quite knows what works and what does not. Companies and researchers take turns introducing new ideas and watching to see what sticks; meanwhile, organizations want AI but cannot agree on how to integrate it. The settled patterns we will eventually take for granted simply do not exist yet.

Most of that broader question is beyond the scope of this post. Here I want to focus on one specific application I have been experimenting with for the past few months: voice.

Voice recognition used to be hard. As a young budding computer scientist in the early 2000s, I understood it as one of those perennially unsolved problems, perpetually five years away. Deep learning changed that, and the change has been so thorough that we barely notice it anymore. Most chatbot subscription services now ship a voice mode in their mobile apps; Claude, my preferred service, is no exception. Curious about its capabilities, I started using it a few months back. I am now hooked.

I have about ten minutes of my daily commute when I am alone in my car. During this time, I can speak freely, and I have found Claude useful as a sounding board for research ideas, paper structures, proposal ideation, and similar puzzles. The format is simple. I start with a meandering monologue about whatever I want to think through, and then I listen to the response.

Project organization matters here, and Claude’s projects (another invention I use religiously) makes it work. I file each conversation in the appropriate project, which means the LLM has access to our prior exchanges (spoken and written) along with whatever Markdown files I have stashed in the project’s knowledge base. This is what turns a series of disconnected car rides into something that accumulates. The conversation on Tuesday morning picks up the threads of Monday’s drive.

The spoken format and the constraints of driving shape the conversation in interesting ways. Obviously, the primary task is driving safely. I cannot scroll back; I cannot edit my prompt; I cannot scan a long response for the part I care about. What I get instead is a linear, oral exchange that resembles thinking aloud to a colleague more than typing at a chatbot. Fortunately, verbalization is marvelously low effort, and the lack of a keyboard removes the temptation to over-structure my thoughts before they have had a chance to develop.

When I get to the office, I switch to the desktop chat interface and ask for a short diary entry summarizing the conversation. The entry goes into the project files along with the raw transcript, adding to the knowledge base. Over time, the project accrues a record of half-formed ideas, dead ends, and the occasional useful insight, all retrievable for the next conversation.

Obviously, there are several potential problems with this approach. Offloading thinking to an LLM is risky for anyone whose business is ideas, which is to say, for us academics. I take care to push back on Claude during these conversations, disagree with its framings, and reject the moments when it tries to smooth my rough thoughts into something more presentable than they deserve. The model has a strong gravitational pull toward consensus and synthesis, and that pull is exactly what you do not want when the whole point is to find that new angle no one has seen yet.

There are also clear limits to what this kind of ideation is good for. I would not use it to draft any real writing. The output of these conversations is raw material: thoughts organized, expanded, occasionally challenged, but never finished. The finishing still happens at a desk, on a keyboard, with the usual editorial discipline. What the car conversations with Claude get me is the upstream stage, the part where you are still trying to figure out what the idea even is.

I suspect HCI and HCAI have a great deal to contribute here. The current voice interface is essentially a port of the text chat with a microphone bolted on; it does not really know that I am driving, or that I am thinking out loud rather than issuing commands, or that the last six exchanges have been one continuous train of thought rather than six independent queries. There is plenty of design space to explore, and driving is only one of the situations where this kind of low-friction verbal ideation could matter. I plan to keep experimenting, and probably writing about it.