The smartphone analogy hits hard. I see this pattern everywhere in AI adoption - people treat transformative tools as glorified autocomplete. The session refresh tip is especially valuable. I've watched teams waste days debugging issues that stemmed from context drift they didn't even know was happening.
Claude Code is a junior developer not a senior. It is a very fast, likely inhumanly fast junior developer that will do exactly what you tell it, and if you tell it to do exactly what needs to be done you will have a good day.
You are right that Claude now auto-compacts context, so you do not need to run /compact for normal, short sessions. Manual compaction is still useful on long or complex tasks though, because it lets you choose the timing and what to preserve instead of having auto-compact kick in mid-task and potentially blur important details.
The example at the start is laughable. Claude never remembers patterns, always recreates, defaults to more fragile methods, ignores instructions and then even says "hey, even when you are explicit I still ignore you 1 out of 4 times."
The only people that think this will work, at present, have never written enterprise software for scale and governance
Agree. Enterprise solutions are a completely different animal and a risky proposition for a Claude Code route. Doable but only afyer maybe years of experience with CC
The smartphone analogy hits hard. I see this pattern everywhere in AI adoption - people treat transformative tools as glorified autocomplete. The session refresh tip is especially valuable. I've watched teams waste days debugging issues that stemmed from context drift they didn't even know was happening.
Thanks for your feedback. I truly agree with that :)
Claude Code is a junior developer not a senior. It is a very fast, likely inhumanly fast junior developer that will do exactly what you tell it, and if you tell it to do exactly what needs to be done you will have a good day.
Its neither, a junior isn't as good at writing code and a senior isn't as bad at logic and understanding the context.
Thank you Nir! Great tips!
You are welcome:)
Great tips. Thanks for sharing. But I was reading context compaction is now automatic in Claude so maybe not required to be done manually?
You are right that Claude now auto-compacts context, so you do not need to run /compact for normal, short sessions. Manual compaction is still useful on long or complex tasks though, because it lets you choose the timing and what to preserve instead of having auto-compact kick in mid-task and potentially blur important details.
The example at the start is laughable. Claude never remembers patterns, always recreates, defaults to more fragile methods, ignores instructions and then even says "hey, even when you are explicit I still ignore you 1 out of 4 times."
The only people that think this will work, at present, have never written enterprise software for scale and governance
Agree. Enterprise solutions are a completely different animal and a risky proposition for a Claude Code route. Doable but only afyer maybe years of experience with CC
Context window limits being the biggest weakness is exactly right. The /compact approach helps but it's reactive. There's a plugin called context-mode that intercepts tool output before it enters the context window and indexes it locally. Means the 200K goes a lot further when raw git and npm output isn't eating it alive. Covered it here: https://reading.sh/how-one-plugin-cuts-claude-codes-context-bloat-by-98-096355e68166?sk=e06168222a38a98d4bf6add2daa10973
Thanks for the feedback! You are welcome:)