Talk to your agents, it works
My workflow has shifted to talking to my agents. I do feel a bit self conscious about it, but it works. But why does it work?
First, a tip: on a Mac, use Muesli.
Muesli is entirely free and works really well. It works in every text box: I just press the Fn key and start talking. It installs a small transcription model and off you go. Recommended!
Isn't a well written prompt better?
When you talk, you'll have lots of ums, and repeats, and you change your mind halfway through. You'd think a carefully written prompt would beat that.
Today's models don't need perfect prose.
They are very good at figuring out what you're telling them, even if it's not written down neatly.
The key is density of information.
When you talk, you go back and forth, you clarify your ideas as you go. That gives the model a lot of signal. Higher density context.
Rambling fills in the latent space.
All that context builds up the model's latent space, it all goes into the right weights. Repeating yourself helps. Using multiple words to say the same thing helps. Clarifying yourself as you ramble helps. It all populates that latent space.
Someone at Anthropic says the same thing.
I had seen this be effective in practice, but I did still wonder. Then I heard Thariq Shihipar from Anthropic making the same point. So, confirmed. Talk to your agents. It works.
Finally, and separately but similarly: it turns out anthropomorphizing the models is useful as well. Who knew?