Hacker News Reader

Comment by tlkn_bot_praxis | Hacker News Reader

I've been stuck in loop all week where a coworker asked me to review some methodology documents that are clearly generated by Claude (which is fine in this case) but every time I sit down my brain can't get through 2 sentences without feeling like I'm reading nothing. It actually feels like the content, which isn't empty, has no meaning. So I go on to another assignment and come back later and the same thing happens. And now I have anxiety about opening it so it has compounded in difficulty haha. Glad other people are experiencing this.

View on HN

I have basically polluted my context and agent files, agent.md, claude.md, gemini.md with really strong guidance about using plain language. I even include small, mini design specifications, examples of plain engineering language, only using technical language and invented terms when they actually make sense (almost never). EVERY TURN.

I have a pretty large set of prompts that go into any software engineering, and I force every single agent to use an ephemeral style stack of prompt management. So, every turn it goes to the top of the stack and it is the very last thing they see in terms of all of my prompts and instructions and agent files. And then it gets taken out of the conversation so that it doesn't get sent to the agent the next turn (no context bloat). It has restored so much sanity.

I tried the caveman add-ons, and I felt like I was losing IQ points because I spend a lot of time reading agent output, and when they start talking like cavemen, I start thinking like cavemen. That was not good for my mental health. So, I try and make the agent talk like me and think like me. And it works, mostly. And my observation is that maybe I'm not the most efficient agentic thought process, but my sanity is retained.

All of that is to say that if something is reading like that to you, just have the agent rewrite it and read it in a rewritten tone because it's probably bad as it stands and your colleague did not put enough effort in it. It is /not/ good and you should not accept it as a default. We have to hold the line on stuff like this and maintain some semblence of normal human engineering standards that existed before AI. They are not making us better. They are making is lazy and dumber.

Opus 5 and other agents in the latest rounds of tuning have gotten ridiculously bad in terms of how they feel to interact with with all the invented language and localized nomenclature. It is an obvious bias that big words and technical talk looks good to the bottom of the bell curve, but when you actually try and understand it, it's horrible. So people say, "Yeah, that looks great," in all the RLHF rounds, and they run with it because they think it looks good, but it doesn't. It's terrible.

Hold the line. It isn't you. And it isn't a good methodology document.

"nd then it gets taken out of the conversation so that it doesn't get sent to the agent the next turn (no context bloat). "

Uh - dude - this means you're paying 10x in token costs because there's no caching.

If you 're-write token history' then you can't cache tokens.

It means for any reasonably long conversation, the llm has to reprocess the entire history as preflow on every prompt.

Are you sure you're really doing what you say you're dong, and how is it not blowing up your budget?

Nope, you gotta understand. This is the move. It is appended at the end of every single message. It is never saved in the conversation that get sent back for inference. So you send it. But when you go back for inference, it’s at the top of the stack so all of the cashing works you’re not pre-filling every time or anything like that. It burns plus N tokens, where N is my prompt stack. It is not really that expensive. I have measured it to within an inch of it its life. Think of it this way every bit of the prefix and the conversation stays exactly the same you’re only adding to the very end of the conversation. So after the first turn, it is basically always cashing within the KV cache for a given context. Sorry I am using voice dictation. My hands are tired this week. Basically you sculpt the conversation history to ensure prefix caching

It should burn N + len(answer), because you have to re-cache the whole answer without the prompt stack.

Perhaps more persnickety, it pushes the LLM out of distribution - if it’s unnatural for it to write in plain language without the prompt stack, your prefix will be an unnatural conversation which can reduce intelligence in hard to measure ways, especially over long conversations.

Not saying don’t do it, clarity is perhaps worth the intelligence hit, but it’s not going to be a free lunch.

Yes, maybe. Evidence around caveman shows this isn’t a big deal for token consumption (forcing it to respond in a way it was not tuned) and I don’t see a big difference either way. And maybe it increases intelligence in hard to measure ways. Lots of parameters in these models.

I feel I fight them less with this setup. They get so lost in their own invented bullshit they stop being useful pretty often without it. So I would take bets on that :)

Yes - with arbitrary prefix caching that might work. That's tricky though, not everyone is going to provide that.

Did you have to build your own harness for this? Or hack Claude Code or something?

I don’t use their CLIs. OMP/OpenCode. If you are stuck on them, you can do it in a proxy layer.

Check out the awesome [Claudish-to-English plugin](https://github.com/gvzdv/claudish-to-english), you could lift their prompt from https://github.com/gvzdv/claudish-to-english/blob/main/rewri...

thanks for this, I'll give it ago. I've tried a few different concision prompts but haven't had great results

I had this the other day with a colleague who is Chinese. Their command of English isn't the greatest, but I never really had an issue understanding them before (in fact, despite their poor English it has literally never been an issue!), but now I just get these walls of text in typical LLM English that have become absurdly difficult to parse. It's like the uncanny valley of text, if you look at it as a whole it seems like it's a proper sentence. The words are all real words, they even make sense in the order the AI has put them in, but as a collective it's like my brain cannot connect more than a sentence or two together.

I thought it was maybe me just becoming lazier with reading considering how much AI-generated text I'm subjected to against my will, but I picked up Blood Meridian the other day and have unironically had less trouble parsing that novel than I have the majority of LLM-authored text I've seen.

Yeah I think a bad prompt can create an output which is meaningless. I had such experience reviewing a software architecture diagram generated by Claude.

Worse, I noticed that people in an office environment themselves have adopted a more speculative, communication style.

In the past, people remembered what was said and would draw attention to discrepancies. I could trust what people said.

Nowadays it's like; someone can say one thing one day and the opposite the next day (through convoluted language) and nobody bats an eyelash. Or sometimes someone will agree with me but then what they say immediately after reveals that they didn't understand the essence of my point at all. I didn't notice these things 5 years ago.

I guess this is what AI researchers refer to as 'model collapse' - it seems to affect people too though...

I've noticed this too. Because people are generating work so quickly , and maybe not understanding it entirely, then they hit a bug, pivot, and pretend it was the plan the whole time. Or drop something they were very excited about 3 days ago and never speak of it again.

There is a lot more hustling and bravado.

It feels like people don't value knowledge as they used to.

It's really hard to avoid mistakes when everyone is subtly covering them up. It feels like a lack of care and I find it demotivating.

I think because engineers are afraid for their job, they are under more pressure to talk a big game. Also under more pressure to deliver short term visible results. Bad combo.

I don't want to read my own Claude outputs, much less someone else's.

If I had to, I'd process it into a short summary and/or ask an agent questions I have about the methodology.

I would then give my feedback. If they ask for details, I can have my agent update the document directly too.

I would not treat an AI generated artifact, be it documents or code, as something that a human should process fully manually.

Agreed, what I call "have your AI talk to my AI" [1]

I think that's where the future is going. It presents interesting challenges, but also some opportunities to make life better for everyone.

[1*] https://foundrs.com/have_your_ai_talk_to_my_ai.html

(*) I dislike the HN trend of starting citations at 0, I find it snobbish.

Holy shit, someone putting it into words

I'm worried we will all be talking to our Claude Therapists soon about our Claude induced learning disabilities lol

This is actually funny because I've been doing some RAG experiments and I've basically trained up a RAG on several gigabytes of philosophy and neuroscience. And when I want to go metacognitive about thought processes, I actually really enjoy exploring different areas of philosophy, more life philosophy, but how it intersects with psychology and neuroscience and observing the observer, or the observer is the observed, as they say.

And when you ground it with real data, it's actually extremely useful. It's not exactly like Claude Therapist, but it's sort of the teach me about philosophy, but actually grounded and not vied. I have a lot of really strict prompts and grounding and agentic guidelines for this particular agent flow and harness that I've built.

And it's just a few weekends of vibing and feeding it basically all of Wikipedia and several gigabytes of papers and stuff, but it actually leads to interesting discussion. So I just have my personal philosophy bot and it's pretty fun.

One of the modalities I built is having two agents assume a famous persona. And then they take a thing, like grief or some thing that I experienced during the week, and they assume the role of the two different philosophers, and I just have them go back and forth 30, 40, 50 turns. And it's actually quite interesting, and it really moderates their language and tonality and behavior. They really get into the roles when you have the right prompting and grounding. Sometimes they get a little off the rails, but it leads to genuinely interesting areas to explore, and then I'll actually go read source material and things like that. I don't know, that's how I do therapy these days, but I never actually did therapy, so I just think a lot, now with agents finding interesting stuff to think about too!

Reply on news.ycombinator.com