Catherine Hicks • 3 months ago
Do not clear the chat when it says so ( use /compact instead)
If you're using command line it will tell you to /clear your chat - if you do that you lose history that makes it a bit of a problem keeping your context if you do that - not impossible to get it back, just a pain when there's an alternative
If you use /compact instead it keeps all your context and doesnt let your conversation get to big.
Log in or sign up for Devpost to join the conversation.
5 comments
Joe Holmes Manager • 3 months ago
Hi Catherine, author of the plugin here. /compact is totally valid! In the earlier days of coding agents, compacting context was pretty unreliable and lossy, but they've improved a lot. I still think there's worth in clearing context between turns as it gives you full control over constructing the context of the agent at each pass, and offloading the particulars to the harness's compaction prompts can feel like a bit of a black box at times. But all the same, totally fine to compact if you get good results.
Catherine Hicks • 3 months ago
Yes but you may lose context as you continue to move forward that may make an impact on a decision further down the road. I started doing what you say and found myself having to repeat myself frequently on things that had been decided, not codified and lost which was getting super frustrating. I may be coming at this from a different angle though cause Im a product designer by trade and having that longer term vision is exactly what I get paid for :)
Joe Holmes Manager • 3 months ago
Glad to hear you've found a strategy that works for you Catherine! Will keep this feedback in mind for v2 :)
Sameera Juman • 3 months ago
You're right, when you start a new session, claude doesn't remember what happened before. A solution to this is creating a Claude.md file by running /init. So when you run some new command, you will have to run a command to modify Claude.md to add that new change that you worked on.
Csaba Toth • about 2 months ago
I guess the plugin author's intention is that context can be confusing sometimes. I over scoped my project so had very long conversations, and sometimes it might be (this needs repeated experiment to prove) better if Claude just works off of what was distilled at the end of the step. If you had to keep repeating yourself that might mean that something (important) was not distilled into the docs. Personally I even installed a memsearch plugin (by Milvus, semantic search).