For me it works at first but Claude models forgets it after some prompts, despite only using like 100k tokens.

Your best bet is to use hooks and inject it after every file edit / response by first running the content through haiku and asking if it is asd 100 ste.

It burns more tokens but is the only way to get tolerable text.

Doesn't it just get attenuated and start ignoring those commands?

The hook sends the text to another agent/context with a request to validate and return a good or bad + reason response. Every request is a fresh context.

https://code.claude.com/docs/en/hooks-guide#agent-based-hook...

Yes but Claude starts ignoring messages when it keeps getting told the same thing over and over.

I don’t think system prompts are particularly reliable way to do much at all. It’s better to put it as a hook after each response, or a skill at least so you can trigger it at will if you don’t want it everytime.

Do you think they're unreliable based on the position in the conversation or other factors?

Anthropic has probably RL’d the system prompt into nothing because of their fear of the user being able to control the model. If it listened to you about the slop language, it might listen to you if you asked it to help you with no-no tasks.

It does work, you however have to put it into every single prompt in which you didn't want a rubbish response

Literally every one, even 1-2 prompts later it starts to go back