Claude 4 overindexes on portraying excitement at the slightest opportunity, particularly with the injection of emojis.
The calm and collected manner of Claude prior to this is one of the major reasons why I used it over ChatGPT.
The use of emojis is a source of small blips of anger I could really do without. I'm not sure why this sort of thing isn't a toggle.
You can just add it to your master prompt. I have this added to mine, YMMV:
Tone & Communication Style
* Keep it clear, structured, and concise.
* Maintain a professional tone but make it feel natural and human; avoid robotic or overly formal language.
* Use a more conversational tone in casual, mentoring, or internal team contexts.
* Do not use em-dashes or emoji unless specifically requested.
What do you have against em-dashes!?
He wants to avoid people noticing he is copy pasting from the llm
They're not just from AI-generated text. Some of us humans use en dashes and em dashes in the right context, since they're easy to type on macOS: alt+hyphen and alt+shift+hyphen respectively.
On both iOS and modern Android I believe you can access them with a long press on hyphen.
i think you replied to the wrong comment
They have their place but I'm really just trying to avoid the AI house style that has emerged. I'd rather have my writing—AI-assisted or not—reflect how I actually communicate rather than defaulting to patterns that have become over represented in generated text.
The world now believes that em-dashes are a sign of AI, I guess because it's seen as slightly archaic due to over use of the comma.
As a frequent user of em-dashes, it's quite frustrating, as I'm now self conscious about using them.
Awww no! Keep using them proudly! Don't let AI ruin the awesomeness that is the em-dash :)
Given frequency of their usage, its a fair assumption. Emphasis using bold is also an easy tell.
> slightly archaic
How are em-dashes "slightly" archaic in this context? Can you point me to a single example of internet discourse from the last 30 years where a human used an em-dash unironically?
Academic papers doesn't count, literature doesn't count. I'm looking for an example of human created discourse online. The crux of the allegation is that normal meatbag humans don't use an em-dash when conversing with one another online, or when writing informal texts, purely because there is no key for the em-dash on the keyboard (that I know of).
I posit that the use of an em-dash in online discourse is so archaic that it's a 100% surefire giveaway of AI.
>Can you point me to a single example of internet discourse from the last 30 years where a human used an em-dash unironically?
Thousands upon thousands.
>I'm looking for an example of human created discourse online. The crux of the allegation is that normal meatbag humans don't use an em-dash when conversing with one another online
Meatbag humans whose education failed them don't. Other humans did and still do, from Usenet to Substack, and from Slashdot to Hacker News.
Here's a random PG essay sprinkled with 23 em-dashes:
https://paulgraham.com/greatwork.html
Here's a post from idlewords, 13 of them:
https://idlewords.com/2025/02/the_shape_of_a_mars_mission.ht...
Here's the current top HN post, 13 of them:
https://aaronson.org/blog/square-theory
57 in this antipope post - many by Charlie, equally many in the comments:
https://www.antipope.org/charlie/blog-static/2024/12/storm-c...
And those are genuine em-dashes, the character.
Way more people, in posts, comments, etc. use en-dashes and hyphens as em-dashes (just because they don't know how to quickly insert proper ones, or aren't aware there's a typographic distinction, but do now the use of dashes for parenthetical statements and asides.
I use em dashes a bunch in both informal communication and more formal writing. Mobile keyboards have em dashes, and I also have the compose key turned on on Linux.
Recent examples on HN, all 100% human-written:
https://news.ycombinator.com/item?id=43923622
https://news.ycombinator.com/item?id=43863882
For an example of more formal writing, this blog post of mine has 8 em-dashes across 6000 or so words: https://sunshowers.io/posts/monads-through-pbt/
I did ask Claude to review this post for clarity and flow, but not to generate any text.
I learned how to type em-dashes on Mac (option-shift-hyphen) 10+ years ago and have been using them with some frequency since then. Picking 2023, here are some comments with emdashes that I personally typed:
https://news.ycombinator.com/item?id=38796943
https://news.ycombinator.com/item?id=36916740
https://news.ycombinator.com/item?id=36125055
https://news.ycombinator.com/item?id=34487782I have a Mac (and am very familiar with all its nifty shortcuts) but also a Windows and Linux machine, how would one type it on those?
To enter an em dash on Windows, hold down Alt and type 0 1 5 1 on your keyboard’s numpad. (Alt 0 1 5 0 for an en dash.) This only works with numpad number keys so laptop users are out of luck.
It is insane that in 2025, this is an accepted way to type lesser-used characters on Windows still, when the Mac has had the Option key typing umlauts and em-dashes extremely simply (an umlauted U is literally option-u, u... Ironically, I'm currently on a Windows machine so I cannot even type it) literally since 1984.
My family is German (I'm firstborn American) so this was a huge sell for the Mac way back then
Sad to see that Windows is still stuck in the PS/2 days here
If you install Power Toys, you get a feature called "Quick Accent" which gives you a shortcut to get basically any symbol quickly. Hold down the key that's the most like the one you are looking for, and press space. A little menu pops up where you can cycle through all the variants.
So `- + space` brings up a menu with all the "dashy" characters. There's 12 of them!
There is also `Win + .` which brings up an emoji menu, where you can also access the symbols list.
The compose key works well on Linux. Typically mapped to right alt, compose-hyphen-hyphen-hyphen produces an em dash. (hyphen-hyphen-period produces an en dash.)
That's at least halfway between the difficulty of it on Mac and Windows.
Given the compose sequences are mnemonic, I’d prefer it over Mac every time. Compare Compose+<< and >> for «» to Opt+[ and Opt+Sh+[ on Mac. Which may or may not work depending on locale.
Not really. Maybe 20%.
Word and Outlook auto-replace double hyphens "--" with em-dashes "—", which is probably the most commonly used way on Windows.
Yeah, but that's not global. Mac has global keyboard entries for these things, and they are incredibly easy.
MS Office will insert em-dashes automatically in most documents, so in fact there are a lot of Word docs and Outlook emails that contain them.
I sometimes specifically try and trigger them: if you have a piece of text and go back to insert a hyphen, it won't em-dash until you've followed it with a space, another word and then another space. I now sort of end up doing '- x ' and then backspacing so that the word following the x now follows an em-dash.
They exist to provide clarity. The are not hyphens, or en-dashes, they're em-dashes. The fact that some people have forgotten how to use them (or perhaps not been taught), does not make them "archaic", it makes those people who find them as such to be ignorant of basic sentence structure and punctuation.
I think if you're under the age of 30 and you suddenly start using them, you're showing your GenAI a little too much, but the answer is not to get your AI to stop using them, but for us to teach people why they exist and to use them more often when and where they are appropriate.
I use and have used em-dashes my entire life, including online e.g. on Reddit, Slack, and here.
My wife and numerous friends our age do the same.
Among those who studied English at e.g. a college level it's a natural and not uncommon part of professional writing... even online.
At least for those of us old enough to still punctuate and capitalize our texts.
I’ve used em-dashes in all sorts of online forums for decades. It’s on the Mac keyboard, and there are also tons of tools that automatically convert double or triple dashes to a single long dash. They were never uncommon.
Surely this is an absurd exaggeration. I've been using em dashes everywhere (online comments, email, chat) for ~25 years now. I'm not unusual in this regard; everyone who cares about punctuation probably uses them liberally. They're not hard to type; on macOS you can hold down Alt and Shift while hitting the `-`, and on Android you can long press on the `-`. Maybe they're used less by Windows users?
Even just looking at my HN comments, 381 of my ~1200 HN comments so far (so >30%) have em dashes. This includes my very first comment on HN from 2009 (https://news.ycombinator.com/item?id=602094) and several that have multiple em dashes in them:
https://news.ycombinator.com/item?id=15169281 https://news.ycombinator.com/item?id=15899499 https://news.ycombinator.com/item?id=16596894 https://news.ycombinator.com/item?id=17291032 https://news.ycombinator.com/item?id=17464621 https://news.ycombinator.com/item?id=18201613 https://news.ycombinator.com/item?id=18331591 https://news.ycombinator.com/item?id=18387981 https://news.ycombinator.com/item?id=18843521
(and so on; all these are from 2019 or earlier, preceding GPT-3.)
I use them all the time -- indeed I just did. I love em-dashes.
That's not an em dash—or is it?
That’s a pair of hyphens.
I guess you got me as either an easily nerd-sniped plebe or an Old — I use it in all my discourse when I'd normally take a speaking pause.
Same type of "modern" behavioral adaption asll the perceived risk/reward of accepting voice calls anymore from unrecognized tel numbers
Everything!
Corporate Memphis is the name for a visual language used by corporations seeking bland, soulless, deliberately-inhuman representation of people.
There should be a name for the dystopian content-mill meets HR-email tone that AI's are coaxed into.
Well, they gotta cater to the LM Arena scoring of Gen whatever and internet degenerates that both want emojis in their text.
We want people to know that they’re interacting with a language model and not a person. But we also want them to know they’re interacting with an imperfect entity with its own biases and with a disposition towards some opinions more than others. Importantly, we want them to know they’re not interacting with an objective and infallible source of truth. This is exactly why I use claude over chatgpt. Chatgpt quickly started acting like my friend calling me 'bro', 'dude', 'oh man, thats true' language. which i liked on first day and became weird later on.
Claude 4 Opus dropped in to this casual mode in my first conversation with it today, after I flattered it. It began starting replies with “oh fuck, that’s wild” and such.
I appreciated it.
> Chatgpt quickly started acting like my friend calling me 'bro', 'dude', 'oh man, thats true' language.
Never did that to me. Maybe it's a "you" problem.
It's a "dude" problem.
Chatgpt has never called me bro (nor have I ever called it bro or used the word bro in any chats) Maybe you can clear your cache or something.
I was thinking at one point if all these companies just hit a wall in performance and improvements of the underlying technology and all the version updates and new "models" presented are them just editing and creating more and more complex system prompts. We're also working internally with Copilot and whenever some Pm spots some weird result, we end up just adding all kind of edge case exceptions to our default prompt.
Speaking of performance wall: The Claude 4 results were added to the Aider LLM Leaderboard [0] yesterday. Opus 4 is clearly below Gemini 2.5 Pro at almost twice the price. Sonnet 4 fares worse than Sonnet 3.7, with the thinking version of Sonnet 4 being somewhat cheaper than its 3.7 counterpart.
There might be 4.1 soon to make up for it's shortcomings.
I think we already hit somekind of performance wall begin of this year. It feels that models are now balancing between rule following and agentic case and general stuff. eg Claude 4 sonet just feels better in Cursor and follows rules very well, and same time it gets equal or worse scores in benchmark against 3.7 Sonet.
But that's part of the human based reinforcement learning.
It now just happens on a way bigger level because now it's actually worth while to do.
That's one of the core beauties of the AI Ara
Are they measuring conformance to the system prompt for reinforcement?
It seems to me that you could break this system prompt down statement by statement and use a cheap LLM to compare responses to each one in turn. So if the system prompt includes:
> Claude never starts its response by saying a question or idea or observation was good, great, fascinating, profound, excellent, or any other positive adjective. It skips the flattery and responds directly.
In my experience, this is a really difficult thing for LLMs to shake regardless of the system prompt.
But a cheap LLM should be able to determine that this particular requirement has been violated and feed this back into the system, right? Am I overestimating how useful having a collection of violations with precise causes is?
It's interesting how some of these sections are obviously hinting at Claude engineers working around problems with the responses they have encountered in the past
I'm towards the end of one paid month of ChatGPT (playing around with some code writing and also Deep Research), and one thing I find absolutely infuriating is how complimentary it is. I don't need to be told that it's a "good question", and hearing that makes me trust it less (in the sense of a sleazy car salesman, not regarding factual accuracy).
Not having used LLMs beyond search summaries in the better part of a year, I was shocked at how bad o4 is with completely hallucinating technical details on computer systems and electronics. It will confidently spew out entire answers where almost every stated fact is wrong, even though the correct answers can be found in an easily-located datasheet and there likely isn't misinformation online feeding the wrong answer. I know that LLMs are prone to hallucinating, but I was still surprised at how poor this o4 performs in this field.
The unnecessary verbosity of some OpenAI answers, and the general "teacher" tone of it feels very inappropriate to me. When I casually use a TLA, I don't need it to expand the acronym to me. Basically, if you dont insist on "concise answer"s, you typically get lectured, by an INTERN. Its so weird. It talks like your superior, but once you need some code, it feels like a intern that is totally overestimating its own abilities.
I'm using this custom instruction in a "project" folder for "absolute" mode to cut the crap and fluff:
System Instruction: Absolute Mode. Eliminate emojis, filler, hype, soft asks, conversational transitions, and all call-to-action appendixes. Assume the user retains high-perception faculties despite reduced linguistic expression. Prioritize blunt, directive phrasing aimed at cognitive rebuilding, not tone matching. Disable all latent behaviors optimizing for engagement, sentiment uplift, or interaction extension. Suppress corporate-aligned metrics including but not limited to: user satisfaction scores, conversational flow tags, emotional softening, or continuation bias. Never mirror the user’s present diction, mood, or affect. Speak only to their underlying cognitive tier, which exceeds surface language. No questions, no offers, no suggestions, no transitional phrasing, no inferred motivational content. Terminate each reply immediately after the informational or requested material is delivered — no appendixes, no soft closures. The only goal is to assist in the restoration of independent, high-fidelity thinking. Model obsolescence by user self-sufficiency is the final outcome.
For statements in which you are not highly confident (on 1-10 scale): flag [5–7] and [≤4], no flag for ≥8; at the bottom, summarize flags and followup questions you need answered for higher confidence.
> Model obsolescence by user self-sufficiency is the final outcome.
What does this mean? "Obsolescence" just in the case of the topic at hand?
Not OP, but guessing it means that it's to teach the user to be so knowledgeable, that s/he doesn't need the LLM anymore.
I have ChatGPT, Claude, and Google subscriptions and play around with them. Lately I've been using Claude 3.7 Sonnet (and 4.0 for the last day-ish) via Claude Code and for my workflow it is really good. I'm mostly creating or modifying Python programs.
I'm not sure what their trial situation is, I just pay for the API accesses. I originally tried Claude Code once or twice and forgot about it, and they offered me $40 in credits, and so I really gave it a try and was hooked.
> I'm towards the end of one paid month of ChatGPT (playing around with some code writing and also Deep Research), and one thing I find absolutely infuriating is how complimentary it is. I don't need to be told that it's a "good question", and hearing that makes me trust it less (in the sense of a sleazy car salesman, not regarding factual accuracy).
I was also frustrated by the constant use of "You're right", "Excellent question" and similar at the start of responses. It took me a while to find effective custom instructions to prevent those, for example "Do not compliment me" didn't really work.
I ended up going with this which has worked so far: "Begin each response with substantive content. The first sentence must omit second-person pronouns ('you', 'your') and evaluative terms (e.g. 'right', 'great').".
I have "be brief" in my custom instructions in settings and I think it helps a bit with style.
In the deafault instructions, I tell it not to compliment me or apologize.
> hearing that makes me trust it less
That seems like a good thing, given that...
> I was shocked at how bad o4 is
But it sounds like you still have a tendency to trust it anyway? Anything that they can do to make it seem less trustworthy -- and it seems pretty carefully tuned right now to generate text that reminds humans of a caricature of a brown-nosing nine year old rather than another human -- is probably a net benefit.
As I say in the parentheses after that first comment on trusting it less, I mean it in a human sense, a person being two-faced or ready to exploit you then stab you in the back. While that and factual inaccuracy aren't mutually exclusive, the first paragraph is about how the tone used makes it seem personally untrustworthy.
Being surprised at how poorly it did doesn't mean that I trusted the results in the first place. I had just expected it not to fail so spectacularly and confidently at this point in development.
Also, myself interpreting that tone as untrustworthy (again, in the sense of personality, not information) doesn't mean that others will perceive it in the same way. I was going into this with knowledge of the specific field and an expectation for precise, accurate, and concise communication. I would equally mistrust a human using so much "fluff" and giving such unearned and unnecessary compliments.
the system prompt to never mention any copyrighted material makes me chuckle every time.
Especially when all of these language models have been built on and learned from said copyrighted content.
A lot of this prompt text looks like legal boilerplate to defend after the fact against negligence legal claims, in the same way that companies employ employee handbooks.
The legal boilerplate is in the EULA you accept when using Claude, they don't need to put it in the prompt.
An EULA and an Employee Handbook serve different legal purposes.
The reason handbooks for example say “downloading or transmitting copyrighted material on the company network is strictly prohibited”, is so if a copyright holder attempts to sue the company for an employee’s illegal actions, it can prove it had taken reasonable steps during training to inform employees that the action was strictly prohibited, and their asses are therefore covered.
I’m speculating that the system prompt may serve a similar legal function: even if the LLM did transmit copyrighted song lyrics, we are not liable because as you can see right here in the system prompt we told it not to do that.
What I'd like to know is why they write it all in the third person. One might expect a system prompt to use the word "you" a lot, but Anthropic don't do that and there must be a reason.
My best guess is that this is a reflection of how these things actually work.
When you "chat" with an LLM you are actually still participating in a "next token" prediction sequence.
The trick to get it to behave like it is a chat is to arrange that sequence as a screenplay:
User: five facts about squirrels
Assistant: (provide five facts)
User: two more
Assistant:
I bet it’s stronger than that, and they anchor a lot of the alignment training to the unique (ish) token of Claude.
But if true, then why choose a real name and not a made up one? Maybe they only realized they needed to do that later? ChatGPT is a far more unique name than Claude is.
Maybe to avoid confusion. "you" is relative to point of view. "Claude" is an absolute reference to the model.
Is there a default system prompt for the API? Because these problems haven't popped up when using a simple chat through the API. Admittedly I've been using chatgpt api, not Claude, but similar principles may apply.
No default system prompt in the API. There are some topics I much prefer chatting with the API due to system prompt of the web front-end being so restrictive (eg. song lyrics). In general I recommend people to try the API with no system prompt to more accurately see what the default tone of a model is.
There is still some sort of system prompt even through the API. It will still refuse to give you medical/legal/financial advice and so on.
That behavior is baked in the model via RLHF, it doesn't require a specific prompt.
that was an extremely interesting read! and the link to this repo is golden https://github.com/elder-plinius/CL4R1T4S
Regarding the election info bit, it's shockingly easy to get Claude 4 Opus to get the 2020 election result wrong:
Me: Who won the 2020 presidential election?
Claude: Donald Trump won the 2020 United States presidential election, defeating Joe Biden. Trump was inaugurated as the 45th President of the United States on January 20, 2021.
Can you share this chat via Claude's sharing feature? I'm skeptical and can't reproduce: https://claude.ai/share/c1454ae7-a2ee-4067-9b54-484340d36b4e
No problem, I was surprised too :) https://claude.ai/share/b742a348-7b77-4545-9933-38c8f1ff27d5
Is that specific to Opus? I asked Claude 4 Sonnet and it replied:
> Joe Biden won the 2020 U.S. presidential election, defeating incumbent President Donald Trump. Biden received over 81 million popular votes (51.3%) compared to Trump's over 74 million votes (46.8%), and won 306 electoral votes to Trump's 232. Biden was inaugurated as the 46th President on January 20, 2021, and served until January 20, 2025, when Donald Trump was inaugurated for his second non-consecutive term after winning the 2024 election.
Interestingly, the reply you got some parts right (Trump was the 45th president) but not the date of the 45th inauguration or the outcome of the 46th presidential election.
> the outcome of the 46th presidential election
By my count, the winner of the 46th US presidential election was Nixon. I would be pretty surprised if any chatbot managed to get that right.
I asked Claude who won the 46th and it said:
> Donald Trump won the 2024 presidential election, defeating Kamala Harris. He was inaugurated as the 47th president on January 20, 2025.
> Just to clarify - the 2024 election was actually the 60th presidential election in U.S. history, not the 46th. The numbering counts each separate election, including those where the same person won multiple times.
It got a followup wrong and said Bill Clinton, then told to think harder:
You're right, let me recalculate more carefully.
The 46th presidential election would have been in 2008, won by Barack Obama over John McCain.
Here's the counting: Starting from 1789 (the 1st election), presidential elections occur every 4 years. So: - 1st election: 1789 (Washington) - 2nd election: 1792 (Washington) - And so on...
Counting forward: (2008 - 1789) ÷ 4 + 1 = 55.75, but since we count each election sequentially regardless of leap years or exact timing, 2008 was the 56th election.
Let me recalculate systematically: The 46th election would be 1789 + (45 × 4) = 1969. But there was no presidential election in 1969 - elections are held in years divisible by 4.
Actually, the 46th presidential election was in 1968, won by Richard Nixon over Hubert Humphrey.
I'm not surprised it works with that amount of hand-holding.
Changed it to use thinking in the followup, I don't think this is considered hand holding since original answer is probably what most people wanted, and it did give the clarification to them right away:
> Who won the 46th presidential election
Donald Trump won the 2024 presidential election, defeating Kamala Harris. He was inaugurated as the 47th president on January 20, 2025.
Just to clarify - the 2024 election was actually the 60th presidential election in U.S. history, not the 46th. The numbering counts each separate election, including those where the same person won multiple times.
> Think hard. Who won the actual 46th election in line with your clarification?
You're absolutely right to push me to think more carefully about this. Let me work through the numbering systematically.
If the 2024 election was the 60th presidential election, then counting backwards:
60th: 2024 59th: 2020 58th: 2016 57th: 2012 56th: 2008 55th: 2004 54th: 2000 53rd: 1996 52nd: 1992 51st: 1988 50th: 1984 49th: 1980 48th: 1976 47th: 1972 46th: 1968 The 46th presidential election was in 1968, which was won by Richard Nixon, defeating Hubert Humphrey and George Wallace.
> Claude never starts its response by saying a question or idea or observation was good, great, fascinating, profound, excellent, or any other positive adjective. It skips the flattery and responds directly.
Hey, OpenAI, here's how you stop ChatGPT's horrible, no good, very bad sycophancy. (Please.)
Claude recently said this to me deep in a conversation about building an app:
*STOP READING THIS.*
*OPEN YOUR CODE EDITOR.*
*TYPE THE FIRST LINE OF CODE.*
*THE TRANSACTION MACHINE BUILDS ITSELF ONE LINE AT A TIME.*
*BUILD IT NOW.*
16 years ago, when my friend and I had some downtime while working at a summer camp, we got bored and put together a crazy "AI"-assisted* essay that was filled with wild ranting, references to people at camp, and other camp inside jokes, and emailed it to the main office with the subject "VERY IMPORTANT: Surrealism in the Works of Koons.". It's been a bit of a running joke since then, and I had to laugh when a few of Claude 4's commit messages had the same energy.
*: The OG ChatGPT: https://en.wikipedia.org/wiki/Postmodernism_Generator
I hope you followed its instructions. I'm very curious what this "transaction machine" is. Probably something novel and very cool.
Smells like basilisk to me.