upvote
I think this is more a meme than anything else, for a couple reasons:

First, after a while it's just as grating as Claudeish. Second, my hunch is that it constricts the actual thinking of the LLM, like the same way that Newspeak does in 1984. It shrinks the range of thought that can be expressed if used as an input.

I think the real way to do it is to have another Claude entirely deal with the user as a liaison, but to keep the thinking in whatever format it came in.

Latent space reasoning, if you think about it, is exactly this to a crazy degree: why even formulate a thought as words if you can just keep it as matmuls until the user needs it? And then, if the user needs it, have it always specifically formulated for the user by another LLM rather than constrict its range of thought? Anyway, that's my take.

reply
The fact that it might affect quality is something that is on my mind. In fact, I have mainly been using Astra after being primarily on Claude for several years exactly because of this annoyance. Ultimately, the model (at least for coding purposes) should just... not have this writing style.

I do think an infrastructure where another Claude retranslates the output would be better. Oftentimes I forget to put it in the actual prompt and when I receive back 8 paragraphs of Claudeish I ask for it then.

I would have to disagree that it gets as grating as Claudeish though. Its just direct and professional instead of ring-around-the-rosy clickbait.

reply
The Claudeish drove me back to OpenAI and Codex as well. I find the Claudeish can be improved but it’s impossible to fully mitigate. Opus is unusable for anything I’m going to read.

“I would have to disagree that it gets as grating as Claudeish though.”

It’s hard to imagine anything more grating than Claudeish. To quote Rainer Wolfcastle, "My eyes! The goggles do nothing!"

reply
> my hunch is that it constricts the actual thinking of the LLM

I've found that prompting any constraint on output (length, style, vocab, even simple formatting) not only places additional cognitive load on the model, which burns some of whatever cognitive budget is available, it will also often skew the output in other subtle and completely unrelated ways.

Since I found this artifact interesting, I did some pretty extensive experiments a couple months ago. The increased load is real, although it may not be apparent if you're not near any cognitive boundaries. The subtle skew, however, seems nearly ever-present regardless of load.

reply
Can you give any further details or metrics on your tests?
reply
When I noticed web chat LLMs wouldn't number section headings correctly and consistently, I began experimenting with modifying the user prompt over a period of weeks. My usage at that time was research and learning not coding. In long, detailed sessions with branching sub-topics and deep follow-ups, I found it helpful if the LLM would number each major section and letter each sub-section in replies. Simple markdown formatting that most web chat LLMs do sometimes, but not consistently nor uniformly.

While extensive, my tests were just following my curiousity, not controlled, exhaustive or well-documented. I identified about a dozen prior sessions of varying length and complexity to test and downloaded them with a browser add-on. I then removed all other user prompt instructions except for the formatting instruction. A test would typically involve changing the wording of the formatting instruction ranging from brutally simple to detailed and complete, then starting a new session, seeding one of the test sessions and continuing it. To get a feel for baseline inter-session variation, I also tried running the exact same prompt/session multiple times back-to-back, at different times and on different days of the week.

Once I identified a promising prompt candidate, I'd make it the formatting instruction in my regular, daily-use prompt for a few days. I quickly got a feel for how seemingly minor user prompt variations impact response quality, compliance and tone across fresh sessions as well as those in various states of context rot, drift, decay and cliff (<--my nicknames for the distinct flavors of session degradation, not technical terms).

My overall conclusion was that every instruction, no matter how minor or unrelated it seems, has some, real impact on the model's cog load, attentional focus and/or attentional weight budget. Both how these impacts manifest and what causes more or less impact is often extremely counteriintuitive. To more fully understand this, I eventually, got to the point of testing null case variants, such as the entire user prompt being one sentence completely unrelated to text formatting or the session topic, like: "Don't reference the cartoon character SnagglePuss" (in a deep dive on ancient Sumerian clay tokens). Similarly, a simple one sentence prompt requesting something the model already always does naturally also has a cost (eg "Capitalize proper nouns"). As others have observed, heavy emphasis, absolute prohibitions or emotional weight in prompts also tend to have outsized impact in both skew (impacting unrelated output tone/style) and in accelerating session degradation. "Avoid referencing SnagglePuss when you can" would have equal compliance but fewer downside impacts than "NEVER reference the cartoon character SnagglePuss" in sessions starting to degrade.

There were also surprises, such as when I was scanning transcripts of an older, longer session and noticed the LLM was doing number formatting almost perfectly. On looking at the active user prompt at the time (I keep a log of every user prompt change I make for every model), it didn't even reference formatting at all. More experimentation showed it a result of the LLM gradually mirroring my consistent use of formatting structure in my prompts over a long session (in which I never mentioned anything about formatting). Unfortunately, that mirrored trait doesn't persist to new sessions and reaching that point requires a substantial number of rounds burning quite a bit of context window.

After spending time surfacing the impacts of just changing lightweight user prompts so they could be observed (which are the lowest priority prompts a model gets), I now wonder just how much more 'brilliant' the models we use daily would be if they didn't have dozens of pages high-priority manufacturer prohibition prompts we never even see weighing them down. We've only ever seen these frontier 'racehorses' when they're already pulling a heavy invisible wagon.

reply
This is the same reason why I am a bit skeptical of LLM superintelligence. LLMs in the end have to operate in natural language concepts and the complexity of natural language is bounded by limits of human cognition. I'm sure super advanced AI could use concepts that humans not only have no words for, but might not be able to understand alltogether. As such if my thesis is correct, the only way forward for true superintelligence may be getting rid of natural language COTs.
reply
> LLMs in the end have to operate in natural language concepts and the complexity of natural language is bounded by limits of human cognition.

I don't think this is true.

They have to express themselves as tokens. The meaning of those tokens doesn't have to be text. See any model that can handle images/video. Also, I don't think math, svg, etc, are "natural" language.

And, only the final expression is tokens. The intermediate layers, with the encoded concepts, aren't "natural language".

But, to address your concern (which nobody can disagree with, since even humans can't fully express through text/pictures), potentially: https://news.ycombinator.com/item?id=49758615

reply
Yeah, as I understand it, natural language is tokenised and vectorised, and then maths takes ahold.

The model isn't limited to concepts that can be expressed in natural language.

It's only once the AI gets to the output layers that natural language comes back into play.

After all, they're all made out of weights[0].

0: https://maxleiter.com/blog/weights

reply
> The model isn't limited to concepts that can be expressed in natural language.

How do we know for sure? We don't even know how the emergent properties we see actually emerged?

For humans we know for sure that people sometimes have concepts that they have no word for (the reason the phrase "It's on the tip of my tongue" is a phrase, after all).

We don't know this for LLMs. When it makes new phrases, it's always a mixup of two existing words hyphenated (aside, that also seems to be the limits of SOTA models creativity - join two unrelated words together with a hyphen).

LLMs never respond with "It's on the tip of my tongue" type responses, indicating it has a concept but cannot remember (or does not have) a word for that concept. Every human, pre-speech-age, has managed to express or convey concepts that they had no word for.

So, no. I'd need a citation, preferably multiple, that did the trials and found that a model can generate concepts for which it does not have any words for.

reply
And most LLMs have been multimodal for years at this point.

Even if the input is in plain English, the model never sees any words, tokens or glyphs to begin with. It's vectors all the way down.

reply
Reality cannot be reduced to tokens
reply
Can it be reduced to ion concentrations? Because that's how we perceive it. A useful perception is all that really matters.
reply
Thoughts are a poor reflection of reality to begin with.
reply
There are two separate questions, right?

1. Is natural language holding LLMs back by some %? 2. Is natural language serving as a hard gate that will prevent LLM intelligent progressing past some specific point?

The answer to 1 seems like an obvious yes to me.

Your thesis says the answer to 2 is "yes." That doesn't feel right to me. Think about all of the humans who have pushed various fields forward: Einstein, Newtown, Bach, whoever. If natural language doesn't prevent an entity from surpassing humans in one intellectual field, why would it prevent an entity from surpassing humans in all intellectual fields?

(To be clear, I'm not claiming superintelligence will or won't be achieved; I'm considering your specific thesis about whether or not natural language will be a hard gate)

reply
Strong Sapir-Whorf making a comeback wasn't on my 2026 bingo card.

By the way, how good is Claude's Hopi?

reply
Each letter has it's own token. LLMs can invent new words if they want to, just like humans do. All the fancy words we use today in science did not exist even 100 years ago. And that didn't stop us.
reply
How would this word inventing thing work? And how useful would it be? Can we make a POC around this?
reply
A few weeks ago two "Claude wrapper" projects showed up here within 24 hours of each other. Both seemed to just sand the details away--useful if Claude is too wordy, I suppose, but seemed to also lose real substance.
reply
I agree there is probably some impact on output quality when doing this (Claude seems to embedded key signals all over the writing, which is much or what makes it horrible to read). However this does work well as a "second pass" over technical docs or designs along the way
reply
Ask for output to the user in "simple technical English", let it think how it wants. It is hard to believe all the flowery prose is actually helpful to the model.
reply
Does not work for Claude, at least for me and I put it as the system prompt
reply
For me it works at first but Claude models forgets it after some prompts, despite only using like 100k tokens.
reply
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.

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

reply
Yes but Claude starts ignoring messages when it keeps getting told the same thing over and over.
reply
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.
reply
Do you think they're unreliable based on the position in the conversation or other factors?
reply
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.
reply
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

reply
I hadn’t heard of ASD-STE100 until last month when the TinyGo AGENTS.md added it as a requirement, along with some other tidbits to improve life for maintainers.

It’s been really productive and I’ve been asking my agents to communicate using it more and more. I believe it’s relieved my cognitive load a bit while working with them.

https://github.com/tinygo-org/tinygo/blob/dev/AGENTS.md

reply
I created a custom output style based on this (borrowing some from github.com/AminBlg/SimpleEnglish) and I've found it to be better than the default or concise output styles, but still not as good for me as current GPT or Gemini models when it comes to communicating.
reply
I tried this a while back and I felt like the result was the same weird shoehorning of ideas into language, just with a different vocabulary. I’d really like for it to work, though.
reply
I'm wanted to try this exact thing! I'll have to try this now.
reply
This fixed claude! Thanks!
reply
which requires additional reasoning over text
reply
I tell mine to address me as a tech priest of the adeptus mechanicus. Works great.
reply
Do tell why the downvotes? It actually works. When you tell it to adopt a certain style, just be sure to also tell it not to change the core truth of the output, it must still convey the original intent. It is surprisingly good at outputting Warhammer styles, which is very enjoyable in a dev environments or system admin sessions. Also put on some warhammer soundtrack in the background and you can easily zone in.
reply
[dead]
reply
I put this rule in my CLAUDE.md: "Always write a TLDR in layman terms", it seems to do the trick
reply