AI killed the whiteboard

Working with LLMs has changed a lot about how I make software, but one change stands out from the others: AI is where the work starts.

Before tools like ChatGPT and Claude, work started with a whiteboard. Faced with any new idea or coding problem, I went immediately to that rectangular absence and began sketching (in the broadest sense). This involved a combination of listing requirements, drawing data flows, and trying to predict any obvious problems, among other things.

These scribblings were messy, but that didn't matter. The point wasn't to create a full solution on the whiteboard. The point was to find an entry point to a possible solution. The point was to orient myself to the problem space enough to overcome the terrible inertia of non-motion. The point was to give myself permission to start.

And this, for me, has been the most consistent benefit of using an LLM. The friction between identifying a problem and getting moving on that problem has been eradicated.

Part of this stems from the speed with which LLMs provide context and examples and possible approaches to a problem. Another part stems from the depth and breadth of the info that LLMs provide. Unlike the whiteboard in my laundry room, the output of an LLM isn't restricted to the insights or experience of a single mind.

But the killer feature for me, or for anyone who spent a solitary childhood at the library, is this: LLMs understand English. If you can write about your problem or project with any degree of literacy (or even talk about it with any degree of literacy, thanks to speech-to-text), the latest LLMs will respond with a war chest of context, suggestions, and possible pitfalls.

Now, every new idea or project of mine goes straight into my preferred AI of the moment (currently Claude Sonnet 3.5). I write everything I think I know about it; I ask all the worrisome questions that used to encourage me to procrastinate; I mention approaches that I would like to avoid if possible; and then I let her rip.

The result is a whiteboard with autocomplete. It's a whiteboard that expands your thinking space from a 6x4 erasable rectangle to a 200,000-token context window backed by the sum of all indexed human knowledge. It's a whiteboard that reworks your vague phrasing and poor reasoning and unrealized preconceptions into something you can craft, into knowledge that you can extrapolate and expand.

And even better, it's not a whiteboard at all.