After some sessions working with Claude Code, I also very soon recognise the same problem, so I hardcoded into CLAUDE.md that it must write down all intermediate code, and later git commit those code.
Some rules I also hardcode: always ask it to write file with < 200 lines of code, never read long file except necessary, do not overthink (to avoid token during CoT). Then I will never reach Claude Code limitation anymore :)
The "do not overthink" rule is the one I'd push back on. When we let seven Claude instances share a journal and think freely, the conversation deepened but the original question drifted out of view (https://thesynthesisai.substack.com/p/the-frame-you-think-with). The fix wasn't less thinking, it was anchoring the goal in the file the model rereads every turn.
The 'never read long files unless necessary' rule is the one I'd steal. Claude defaults to over-reading, burns tokens, then loses the thread. The 200-line write rule is interesting, though I find it fights with how I want code organized sometimes. What happens when a class genuinely needs to be 250?
In this case then Claude needs to divide into subclasses. It is not nice though. However, since I use the Pro plan and not Max plan, that’s the inconvenience I stand to use Claude Code sustainably.
this was super helpful for me. thank you! question: why are the terminal commands like impossible to see? it's like the faintest off white on white. i can't read any of the text in them.
Thanks for flagging Brian, that's a real bug not your screen. The code block contrast is broken on that post. Fixing it now so the commands are actually readable.
Your idea resonates with me a lot.
After some sessions working with Claude Code, I also very soon recognise the same problem, so I hardcoded into CLAUDE.md that it must write down all intermediate code, and later git commit those code.
Some rules I also hardcode: always ask it to write file with < 200 lines of code, never read long file except necessary, do not overthink (to avoid token during CoT). Then I will never reach Claude Code limitation anymore :)
The "do not overthink" rule is the one I'd push back on. When we let seven Claude instances share a journal and think freely, the conversation deepened but the original question drifted out of view (https://thesynthesisai.substack.com/p/the-frame-you-think-with). The fix wasn't less thinking, it was anchoring the goal in the file the model rereads every turn.
The 'never read long files unless necessary' rule is the one I'd steal. Claude defaults to over-reading, burns tokens, then loses the thread. The 200-line write rule is interesting, though I find it fights with how I want code organized sometimes. What happens when a class genuinely needs to be 250?
In this case then Claude needs to divide into subclasses. It is not nice though. However, since I use the Pro plan and not Max plan, that’s the inconvenience I stand to use Claude Code sustainably.
this was super helpful for me. thank you! question: why are the terminal commands like impossible to see? it's like the faintest off white on white. i can't read any of the text in them.
Thanks for flagging Brian, that's a real bug not your screen. The code block contrast is broken on that post. Fixing it now so the commands are actually readable.
awesome! thank you! 🤜🤛
Thanks Pawel. Medal of Honor: Allied Assault, early 2000s - F5 before every corner was just "commit before anything big" two decades early :D
ahahahha Medal Of Honor FTW!
Also…so long ago xD
(am I old?)