ChatGPT Voice is easily the future of work for the next 6 months. I’m able to type about 98 WPM on my MacBook Air with no warm up and usually break 110 when on my mechanical keyboard. The input from my brain to the model does not seem to significantly speed up when using ChatGPT voice unless I am rambling a large project that will need several different features developed.
For the input method, my recommendation is to ramble as much as you can possibly think of, and at the end ask for separate threads with worktrees running either Sol extra high or Luna Max depending on your tasks and have those trees notify the voice thread. This works really well if you want to iterate or change direction on an idea and Voice is really good at figuring out how many threads you need and which should be doing what.
The real power comes during the iterations and review. I can listen much faster than I can read. Not only that, Voice mode can look at the diff to find the dependencies or how a function is used much faster than I can read all the code and construct the architecture in my brain. Looking at the diff while Voice explains it has probably 100x’ed the speed at which I can review an implementation. With a strong CI/CD test suite and linting I don’t need to pay attention to the static analysis as much and can focus on the direction and architecture.
A lesson I learned though is that you need to make a Scheduled automation that will run weekly looking for old worktrees (if you’re using GitHub this is easily identified by PRs that are merged or stale) and then cleaning up. If not you can end up with some very large worktrees with virtual env’s or dependencies that have been downloaded multiple times to run your product or test suite. Luna Max is great for this cleanup.