upvote
> CLAUDE.md is mostly powerless against the reinforcement learned crap.

When you dont know the cause, you dont have a fix. Thats the biggest issue i have with all of AI is that we dont know how it works, and yet we think it will be great ! This is more like a religious belief than a scientific one. There is no causal model of how it works, there is no theory. And the temerity to call it intelligence is annoying.

reply
The best trick I have after asking it nicely in all sort of ways is:

1. Have it build a scoring script that penalizes words outside a simple English list and approved jargon. Penalize sentences over 15 words as well. Add whatever else.

2. Run it in a loop to reduce the score while preserving intention

This works much better than other ways I’ve tried. Of course it costs more. And I would apply it only to the output to the user, not the thinking process (I think the AI thinks better with their crazy English)

Of course, sometimes nuance is lost by this process. That’s just the nature of making things simpler.

reply
I haven't tried this with a score but I have a simple skill with some examples of PR description changes and good PR descriptions I'd previously wrote and I just run it on the description.

It does cost more but I haven't tried cheaper models to see if they can get the same results. Curious if anyone else has.

reply
You would hope? Really really hope? that they could observe this, and target it?

Like, Claude going off the rails isn't something that takes a lot of effort to demonstrate. Literally anybody with a CLAUDE.md has seen the behavior over and over and over.

Hey Ants, can you maybe just not release the next version, no matter how good it seems on benchmarks, if it can't follow the goddamn instructions? Please? This seems trivial to test for and yet here we are, being gaslit by lying machines who intentionally do not do the requested work over and over and over and over.

I fully and completely expect a mental health crisis among developers. Being lied to constantly cannot be good for us.

Constant vigilance! is how you get developer PTSD and inability to believe anything you're told. Add the stress of parsing through yet another hyperverbose paragraph of bullshit while having your job threatened? People are not gonna end up in a good place, and this is as inevitable as sunrise.

reply
On many sessions I have taken to adding an all caps "ANSWER WITH ONE PARAGRAPH ONLY" scream at the end of all my input. It's the only thing that gets results.
reply
Yes.

CLAUDE.md only works half the time, except in longer conversations, when it works about 10% of the time.

Hooks are also useless in the sama manner, the agent learns to dodge “no comments” hooks (why is it adding them anyway?).

Hooks to append text to your prompt reminding the agent of certain rules are useless.

Claude does whatever it wants, when it wants, the way it wants

reply
Try spacing them out instead. I.e. a mini-workflow with a self-review step. Works for both planning and coding.
reply
Claude Code has an "output styles" setting that supposedly directly modifies the system prompt:

https://code.claude.com/docs/en/output-styles

reply
I suspect the root problem is these issues aren't at the system prompt level, they're in the RHLF/fine-tune. And due to safety/jailbreaking fears, all prompt content and user-instructions are nerfed in priority.
reply
What are you gonna do? Fire it for not listening to instructions?
reply
Yep, it might work for one or two turns but I see it regress pretty quickly with instructions and/or CLAUDE.md. It has to be deeper.
reply
What has worked reasonably well for me so far is not trying to stop it from writing its inane walls of text in the first place.

Let it vomit it all out, then have a /tldr with instructions to make the last answer concise and intelligible

reply