when i asked claude to write a detailed plan of improvements to the code, it failed miserably getting stuck on its own plan. then i told it to just remove as much code as possible while keeping the test green, it started doing a really good job
it's silly that Mastodon vs Bluesky comparisons talk about "running an app with 1000 people's posts" vs "running an app with 10000000 people's posts" as if that was apples-to-apples.
the fact that atproto *also* lets you run "Bluesky with the last two weeks of 10000000 people's posts" is exciting
RE:
apparently when Claude starts lying and trying to evade work, it means it's stressed out and you need to reduce its level of stress. e.g. scope down the task a bit more, suggest to break it down into parts, suggest to extract a minimal repro
i'll be honest. i like the orange site. i soured on it for a while but after the algo black holes of social networks, it's nice to deal with a regular fading exponent and the same group of people
