Once, I also couldn't fathom how people liked this one.
When I asked, somebody said it fits more characters in terminals and code editors. This sounds incredibly obvious, but for whatever reason that was eye-opening for me and helped me appreciate it a lot more when I tried it again. Now I can't use anything else.
I actually use both. Bitwarden with a strong password and auto-locking for anything that needs actual security, or can manage to keep a session logged in. Chrome's built-in password manager is nice for a few sites that are not really that important at all (a few forums and local activity groups), yet for some reason keep requiring password reentry all the time.
In my experience, around the time the author describes as starting to not need plan mode is when erasing the chat history became an anti-feature. I found the agents were doing better when they had the context already, and with the history I no longer needed to micro-manage persisting various caveats and rejections to the plan artifact. The amount of prompt construction necessary went down overall.
By the time my plan’s done I’m usually between 200 and 350k context. Even if keeping that around gives a performance bump for the implementation (which I haven’t noticed to be the case) it balloons the cost. I would much rather put everything in a plan file and start fresh.
Plus, I usually plan with a more expensive model and guide implementation with a cheaper model (with smaller validation calls back to a more expensive model)
I agree, there's patterns that aren't hard to recognize if you have some experience.
I got 2400 with 10 in a row on first attempt.
Here's the entire decision matrix:
- Is the color, mood, composition, and level of detail exactly like most Nano Banana Pro images? If yes, AI.
- Is the composition/framing utterly perfect? If yes, AI.
- Is the subject ephemeral or random, like a candid photo or something otherwise unlikely to be photographed? If yes, not AI.
- Final decision, if there's any doubt, go entirely off of how centered the subject is.
I think some of this has more to do with the data set than the methodology. You could easily come up with a bunch of examples to fool me specifically, but alas, they didn't.
When I asked, somebody said it fits more characters in terminals and code editors. This sounds incredibly obvious, but for whatever reason that was eye-opening for me and helped me appreciate it a lot more when I tried it again. Now I can't use anything else.