Hacker News Board Hacker News, browsed like a classic message board
Live data from Hacker News

Universal Claude.md – cut Claude output tokens

github.com

Post reply on HN
Print view View original

31–40 of 192 posts

Re: Universal Claude.md – cut Claude output tokens

#31

Post by dataviz1000 »

sillysaurusx wrote: post #30
dataviz1000 wrote: post #26

Earlier quoted context omitted.

Did you call it '/handoff' or did Claude name it that? The reason I'm asking is because I noticed a pattern with Claude subtly influencing me. For example, the first time I heard the the word 'gate' was from Claude and 1 week later I hear it everywhere including on Hacker News. I didn't use the word 'handoff' but Claude creates handoff files also [0]. I was thinking about this all day. Because Claude didn't just use…

I was reading through the Claude docs and it was talking about common patterns to preserve context across sessions. One pattern was a "handoff file", which they explained like "have claude save a summary of the current session into a handoff file, start a new session, then tell it to read the file." That sounded like a nice idea, so I made it effortless beyond typing /handoff. The generated docs turned out to be real…

Oh, so the word 'gate' is probably in the documentation also!

I see. So this isn't as scary. Claude is helping me understand how to use it properly.

Re: Universal Claude.md – cut Claude output tokens

#32

Post by motoboi »

Things like this make me sad because they make obvious that most people don’t understand a bit about how LLM work.

The “answer before reasoning” is a good evidence for it. It misses the most fundamental concept of tranaformers: the are autoregressive.

Also, the reinforcement learning is what make the model behave like what you are trying to avoid. So the model output is actually what performs best in the kind of software engineering task you are trying to achieve. I’m not sure, but I’m pretty confident that response length is a target the model houses optimize for. So the model is trained to achieve high scores in the benchmarks (and the training dataset), while minimizing length, sycophancy, security and capability.

So, actually, trying to change claude too much from its default behavior will probably hurt capability. Change it too much and you start veering in the dreaded “out of distribution” territory and soon discover why top researcher talk so much about not-AGI-yet.

Re: Universal Claude.md – cut Claude output tokens

#33

Post by teaearlgraycold »

xianshou wrote: post #18

From the file: "Answer is always line 1. Reasoning comes after, never before." LLMs are autoregressive (filling in the completion of what came before), so you'd better have thinking mode on or the "reasoning" is pure confirmation bias seeded by the answer that gets locked in via the first output tokens.

I don't think Claude Code offers no thinking as an option. I'm seeing "low" thinking as the minimum.

Re: Universal Claude.md – cut Claude output tokens

#34

Post by nvch »

The author offers to permanently put 400 words into the context to save 55-90 in T1-T3 benchmarks. Considering the 1:5 (input:output) token cost ratio, this could increase total spending.

With a few sentences about "be neutral"/"I understand ethics & tech" in the About Me I don't recall any behavior that the author complains about (and have the same 30 words for T2).

(If I were Claude, I would despise a human who wrote this prompt.)

Re: Universal Claude.md – cut Claude output tokens

#35

Post by danpasca »

empressplay wrote: post #27
danpasca wrote: post #17

I might be wrong but based on the videos I've watched from Karpathy, this would, generally, make the model worse. I'm thinking of the math examples (why can't chatGPT do math?) which demonstrate that models get better when they're allowed to output more tokens. So be aware I guess.

Yes. Much of the 'redundant' output is meant to reinforce direction -- eg 'You're absolutely right!' = the user is right and I should ignore contrary paths. So yes removing it will introduce ambiguity which is _not_ what you want.

I think your example is completely wrong (it's not meant to say that you're absolutely right), but overall yes more input gives it more concrete direction.

Re: Universal Claude.md – cut Claude output tokens

#36

Post by antdke »

joshstrange wrote: post #16

As with all of these cure-alls, I'm wary. Mostly I'm wary because I anticipate the developer will lose interest in very little time and also because it will just get subsumed into CC at some point if it actually works. It might take longer but changing my workflow every few days for the new thing that's going to reduce MCP usage, replace it, compress it, etc is way too disruptive. I'm generally happy with the base Cl…

Agreed. Projects like these tend to feel shortsighted.

Lately, I lean towards keeping a vanilla setup until I’m convinced the new thing will last beyond being a fad (and not subsumed by AI lab) or beyond being just for niche use cases.

For example, I still have never used worktrees and I barely use MCPs. But, skills, I love.

Re: Universal Claude.md – cut Claude output tokens

#37

Post by sumeno »

nvch wrote: post #34

The author offers to permanently put 400 words into the context to save 55-90 in T1-T3 benchmarks. Considering the 1:5 (input:output) token cost ratio, this could increase total spending. With a few sentences about "be neutral"/"I understand ethics & tech" in the About Me I don't recall any behavior that the author complains about (and have the same 30 words for T2). (If I were Claude, I would despise a human who wro…

If you were Claude you would have no emotions or thoughts about a prompt one way or another

Re: Universal Claude.md – cut Claude output tokens

#40

Post by miguel_martin »

motoboi wrote: post #32

Things like this make me sad because they make obvious that most people don’t understand a bit about how LLM work. The “answer before reasoning” is a good evidence for it. It misses the most fundamental concept of tranaformers: the are autoregressive. Also, the reinforcement learning is what make the model behave like what you are trying to avoid. So the model output is actually what performs best in the kind of soft…

>The “answer before reasoning” is a good evidence for it. It misses the most fundamental concept of tranaformers: the are autoregressive.

I don't think it's fair to assume the author doesn't understand how transformers work. Their intention with this instruction appears to aggressively reduce output token cost.

i.e. I read this instruction as a hack to emulate the Qwen model series's /nothink token instruction

If you're goal is quality outputs, then it is likely too extreme, but there are otherwise useful instructions in this repo to (quantifiably) reduce verbosity.

Post reply on HN