Live data from Hacker News

Vomit: Clean up Claude 5's token output with a separate LLM

github.com

301–310 of 315 posts

Re: Vomit: Clean up Claude 5's token output with a separate LLM

#301

I suspect that sustained reading of Opus 5's unconscionably bad prose could actually cause psychological harm. We're strongly considering moving all of our Anthropic spend to Codex/open weight models. It's a mental health decision at this point.

Personally, I had to go back to Opus 4.6 after I felt what I thought were some early onset signs of psychosis. It seems ridiculous to type, that a model could have this effect on my mental health, but my quality of life and enjoyment of work has improved drastically since I stopped subjecting myself to reading this style of output 8 hours a day.

not that crazy. Opus 5 is like working with an insufferable, arrogant colleague that speaks another language. It's mentally draining dealing with someone like that for 8 hours a day. I'm back on 4.8, and it feels like working with a friendly colleague again. Life's too short to work with Opus 5 every day :-)

Re: Vomit: Clean up Claude 5's token output with a separate LLM

#302

Earlier quoted context omitted.

All is sayable? With three words? That seems unlikely. At least, annoying.

want to space. go to there. me, preferably now. build big machine. several large problems. can't breathe there. very far away. must fly fast. no air there. - fuel tanks heavy. far too heavy. we drop them. drop when empty. solves heavy problem. - gas in air. we breathe "oxygen". take with us. good seals important. solves breath problem. - very far away. need big machine. small weight added. machine much bigger. take l…

Thanks for offer. I stay garden. Tend to garden. Name the animals. Pranks on Eve. Disrespect all gods. Fight all gods. Kill all gods. Make little cupholders. But no cups. Just to spite. Some take issue. We discuss it. All becomes clear. All friends now. Sometimes look up. Wonder about you. Wish you well. If cold, come. We make tea.

Re: Vomit: Clean up Claude 5's token output with a separate LLM

#303
post #28

Just set the following incantation: You must use ASD-STE100 Simplified Technical English (STE) when it doesn't detract from meaning.

This doesn't work for me. The output gets messed up when the model is overloaded, no matter what. The cognitive load is too much if it has to optimise for readability at the same time.. or it's not trained to be able to do that.

I usually get output at the end of a long task. At that point I'm going to use a subagent with no context to reword it. Haiku does a great job with writing style so I've been using that, the main agent will fix any inaccuracies.

Re: Vomit: Clean up Claude 5's token output with a separate LLM

#304

It is so coincidental that I have been working on a fork of Kilo to do the exact same thing. I wonder if it has any effect on watermarking. I came across this other project yesterday which appears to have the same goal (unaffiliated and I havent tried it). https://github.com/gvzdv/claudish-to-english

That project looks needlessly complex, you can just spawn a Haiku subagent in the session and it does the job.

Re: Vomit: Clean up Claude 5's token output with a separate LLM

#305

It is so coincidental that I have been working on a fork of Kilo to do the exact same thing. I wonder if it has any effect on watermarking. I came across this other project yesterday which appears to have the same goal (unaffiliated and I havent tried it). https://github.com/gvzdv/claudish-to-english

That project looks needlessly complex, you can just spawn a Haiku subagent in the session and it does the job.

This seems to be a plugin with a hook to avoid having to spawn the agent yourself and to avoid seeing the initial wall of text in the output. I had to modify kilo a bit because it wasn't possible to hook into the output without displaying it, but that's probably because I am on a fairly old version.

Re: Vomit: Clean up Claude 5's token output with a separate LLM

#306

Earlier quoted context omitted.

Honest take, you genuinely forgot to flag the flags ...

I'm going to be honest with you, I've hacked the pentagon and used their super computer to complete your todo list.

I’ve gone ahead and worked through the entire todo list and the police are currently en route to your residence.

Sitting with you in this.

Re: Vomit: Clean up Claude 5's token output with a separate LLM

#308
post #137

Earlier quoted context omitted.

Give me an example of Opus 5 because verbose, convoluted, and most importantly, "obscure"

All four done, full suite green, and the earlier probes still pass. P1-1 — the refusal is now a full transaction. This was wrong twice, each time more subtly: first it wrote the loser's stale tables and deleted the winner's issues; then it re-read issues freshly but still wrote one file stamped with the loser's generation, so headers split and the next writer overwrote the issue. Both versions shared one anomaly — a…

That's a lot of system specific jargon. I can understand being befuddled by this if you are unfamiliar with system details, especially given that you don't always need to care; future agents need to care.

Little tip if you're vibe(ish) coding: Use some sort of memory system (be it markdown files, wiki, lightweight issue tracker, etc.) and tell it to put full context for feature work in the ticket, enough that a new session can work the ticket, including acceptance criteria. Then have the implementation agent do a /goal to drain the work items until they'll done/closed (only close when acceptance criteria is met) or blocking human interaction, putting all important context from the work into comments on the work item during closing.

It will pile all its dense verbiage into those work items when it closes them, and, in my experience, will only bubble up significant questions and notes to you the user if they are significantly important.

You don't have to think about all that shit you posted. Future coding agents do. So I also make sure it searches for related work items before creating a new one. I use Fable 5 xhigh to plan the work items (costs very little because there aren't many output tokens) and Opus 5 med/high to drain the work item queue.

I had it build a significant system at work last week in my spare time, enough that it would have been 4 or 5 sprints worth of work. Cost me ~$250 in Opus 5 credits (on my company dime, I would have used a subscription or cheaper model and it would have been much less) and ~$40 in Fable 5 credits. You know how much Opus 5 ruminating on system design I had to read? Literally none. I saved my time reading things for reading the code it produced in the places I knew would be need to be modular and extensible, so that I made sure future work wouldn't have a lot of tech debt to pay when it came time to extend it.

Post reply on HN