Claude: System Prompts
platform.claude.com
platform.claude.com
For example here's what changed between Opus 4.8 and Opus 5: https://github.com/simonw/research/commit/a2de185cc367eb66c2...
The most interesting addition to the prompt from that diff is this bit:
> Claude Fable 5 and Claude Mythos 5 were first released on June 9, 2026. On June 12, 2026, Anthropic suspended access to both models to comply with U.S. Department of Commerce export controls; the Department lifted those controls on June 30, 2026, and Anthropic restored access on July 1, 2026 (Anthropic's statement: [https://www.anthropic.com/news/fable-mythos-access](https://www.anthropic.com/news/fable-mythos-access)). These events are after Claude's training-data cutoff, so Claude knows about them only from this notice. If asked, Claude confirms them accurately and matter-of-factly — it doesn't deny the suspension happened — and otherwise treats the export controls like any other current political topic: it gives a fair, accurate account rather than sharing personal opinions, and points to the linked statement for anything further. Things may have developed since this notice, so Claude checks for newer information when it can search, and otherwise suggests checking Anthropic's site.
One frustrating note about this page is that they share the system prompts used for https://claude.ai and the Claude mobile apps regular chat, but they omit the tool definitions. Those are much more interesting if you want to understand what Claude can actually do for you. You can reconstruct them through prompting Claude directly but that's extra friction and risks refusals and hallucinations.
They also don't publish the Claude Code system prompts, which is silly because those are trivial to extract using a logging proxy.
I don't really think 1 line in lengthy system prompt affects things that much, it'd just be an amusing form of emergent behaviour where we now have to massage the ego of something with no id.
Start short or empty and watch where it makes mistakes then just keep tuning it so they're less frequent. That works for me.
Cant it spin up a webbrowser in the background and go to claude.ai and play with the sibling models and "find out" about it rank? :-D
at the least it takes away thousands of tokens in the most important part of the context window (!)
also see the comment by comboy on contradictions not helping performance
the system prompt is the most important part of the instruction you can give the model
it comes before everything else + the model is trained to pay extra attention to it
edit: that's also why in smol (minimalist agent harness) there currently is no system prompt at all (you can add one easily if you want to though)
https://github.com/smol-env/smol
the context window is precious
it should be filled with your task and helpful context for that task
afaiu they have other systems for denying and re-routing requests
But how do you start a Claude Code session without a system prompt?
Otherwise DANmode and similar jailbreaks would still be as easily accessible as they were at the beginning.
Any mention in the system prompt is mostly defense in depth, and to make refusals more graceful.
If the prompt were the only protection, it would be extremely easy to produce illegal content after a long session.
Last time I tried these were trivial to get by just putting a man-in-the-middle proxy between Claude Code and Anthopic. You get both the system prompt, and all of the tool definitions. So there's no need to go about it in such a roundabout way.
From my original comment:
> They also don't publish the Claude Code system prompts, which is silly because those are trivial to extract using a logging proxy.
For those interested in the Claude Code tool descriptions, they're collected here: https://github.com/Piebald-AI/claude-code-system-prompts/tre...
As for the tool implementations, there are still remnants from the Claude Code source leak visible online, with a bit of research.
Few days back, I posted an article[1] that was about how AI threatens natural resources for billions. This was from United Nations and it was flagged. I did not think much about it until I saw two other stories [2] & [3] today that were doing fairly good on front page but they suddenly disappeared. They are not even on 2nd or 3rd page. I have seen this happening at other times as well but did not document it. Just thought you all should know about this.
I was going to create Tell HN thread but I thought the same would happen with it too. I am pretty sure this thread is not going anywhere so I'm posting my concern here.
[1]: https://news.ycombinator.com/item?id=49290062
It reminds me to how various political movements will complain about being unfairly censored, pointing at their posts being disproportionately removed as evidence of this, then you look at said posts, and discover that they're simply disproportionately questionable in the first place.
There's definitely merit to monitoring something like this, so I do appreciate you surfacing this here, but there's also definitely a wheat and a chaff to this, and so based on just this much I have to disagree.
A sibling comment mentioned a few details already, but there's a good amount of information out there about how HN's post ranking system and moderation works, that'd probably be good to also consider. Maybe reaching out to the mods would also be helpful in the way of this.
To give you an anecdotal example, if I see a post mentioning how LLMs are "just next token predictors", I'm basically flagging that by reflex at this point. Not because it'd be literally untrue, but because it's asinine overall. But you won't be able to infer this from data, only the fact that a post using "AI-critical language" was flagged.
Or there was another post about how Ireland's electricity use is so-and-so % data center driven, further suggesting that this is trending up. This was not true, and the article was further horribly unhelpful in actually putting this fact into context, or properly conveying the trends. I think I ended up flagging that one as a result, after posting - what I thought - was a lot fairer picture (and even that was awfully lacking in context). Once again, an "AI-critical" post which on the face of it would have been simply censored if enough flags gathered.
There's also the mundane human angle to this, where people enthusiastic about <thing> won't necessarily be the most receptive to criticism to it, and will be more likely to try and pick that criticism apart. Gotta match the audience on some level.
You are doing a disservice as people going through AI psychosis should hear that LLMs are just calculators.
Not any more than I do by not going around telling the depressed to simply chin up, or telling people who are frustrated with Trump that they're simply sick with TDS.
There's no world where instigating others through snide insinuations works out in one's favor. You may feel extremely justified in acting like this, but I don't think you'd appreciate if you were met the same way in other subjects either. Unfair as it is, an escalation is an escalation, no matter the context or the justification. It's always a lose-lose.
You can remind people that the conversations are simulated and that LLMs are just programs without reductively handwaving how they work. You can express frustration with the overenthusiasm around LLMs without accusing others of mental illness. I guarantee you that it lands a lot better and provides a significantly better "service" than shouting ragebait into the void ever does. Give others a chance to be better than what you think of them. You may be surprised.
Consider post [3] that the parent commenter cited as an example for a post that had a "negative connotation to AI". In reality, it doesn't take much reading to confirm it was anything but. But even if we pretend that it was, and assume it was some hit piece about how "all the people who are now willingly having their brain fried will have to wake up to a grim reality", it doesn't take much to see that token frugality and business justification of AI-use is very much in the interest of those who do feel enthused about LLMs. So there'd be no reason to present the topic so maliciously; on the contrary, the author could tailor the article around this other perspective, and they'd reach a new audience with the same fundamental message, and have them cooperate too.
Tangentially related in its principle: https://www.youtube.com/watch?v=s1EVk7k9S7Q
Are we reading the same site? There is constant anti-AI content on the front page.
I know you're upset that your submission got flagged, but as most of the comments pointed out it wasn't even a well-argued piece. It got flagged because users here expect to read reasonable arguments, but the commenters called out real problems with the article and the arguments it was failing to make.
I would take it as feedback about what types of articles aren't welcome by the users here, not an indictment of the specific topic you submitted.
The first example was flagged by users. It fits the pattern of other political clickbait stories. The top comment is calling out problems with it. This type of story pops up and gets flagged all the time on different topics.
Some people assume a conspiracy or moderation misbehavior, but when most of the comments in the thread are people calling out obvious problems with the article it leads to a lot of users clicking the flag button. Articles with poor logic or tortured claims don't last long here.
The second one is an Ask HN on a contentious topic with more comments than upvotes. There’s an automatic filter on this website designed to detect flame wars and I suspect it down ranks threads that aren’t getting many upvotes but are attracting a lot of comments. Happens to many Ask HN threads.
The third one doesn't even strike me as anti-AI. I don't know why you included it as an example of an anti-AI agenda because it's still about a future where everyone is using AI. It has other problems though because it's willfully ignoring the fact that inference is getting cheaper at a fast rate. It probably got dropped from the front page because the ratio of comments to upvotes was bad, like the other story.
There isn’t a conspiracy theory to be found in these examples. This is just what happens to tired topics on this site.
Anti-AI topics are on the front page all the time. I think that story you tried to post was just a badly written anger bait piece, it got called out in the comments, and people started flagging it.
Read the comments. People were actually reading the topic and calling it out. This gets topics flagged.
It’s not coordinated interest groups conspiring to remove stories.
> Give me a break, Mr. "nothing to see here".
Okay, Mr. “I just created an alt account for this comment”
False. I occasionally flag items, never as part of nor acting on behalf of any particular interest group.
That’s what they mean. It’s too easy to flag something here, as it is too easy to downvote.
He didn’t say that at all. He said there was bias towards removing negative AI posts. He showed no knowledge of the flagging system and that it could just be working as intended.
If anything, they don't get flagged nearly consistently enough.
See anything below too?
https://news.social-protocols.org/stats?id=49290062
https://news.social-protocols.org/stats?id=49318906
https://news.social-protocols.org/stats?id=49319582
Edit: previous sibling comments explain pretty well + imagine moving the needle on AI on HN with negative coverage of it!
The article rests on the claim that water usage of data centers on continent X threaten water availability for humans on continent Y.
I hope you can see how self-evidently illogical that is. The article tries to bend logic into a narrative it is trying to push.
> This was from United Nations and it was flagged.
Indeed! It exposes a level of lack of rigor and critical thinking which is astounding - this is meant to be an organization that thinks clearly, which it clearly does not.
The actual problem is making the desalination tech itself cheap and easy to maintain so poor countries can afford it. Right now it's only a large scale thing in relatively rich and tech advanced states like Saudi or Israel.
Which is all to say in addition to the crazy energy demands of desalinization, there are also problems that dwarf those.
The people in continent Y are likely to use extreme violence to stop this from happening, like we see today in the Mediterranean sea where thousands of people are murdered every year.
I don't want to start too much of an off-topic thread within an already off-topic thread, but this isn't a particularly charitable reading. The article itself is mostly laying out information around data centres' resource usage with little prediction for how things would play out. I think statements like:
> The water footprint is equivalent to the minimum annual domestic water needs of roughly 500,000 people in Sub-Saharan Africa
Are clearly just intended as illustrative, rather than a claim that 500,000 people in Sub-Sarah Africa will go without water as a result of AI data centres.
I think the articles head isn't particular good, but it's a fairly moderate take and it's closing message isn't an "AI dooms natural resources" one at all:
> AI within planetary limits is achievable. The report’s central argument is constructive. Capability and stewardship can grow together, but only with measurement, transparency, and shared responsibility across the ecosystem.
At any rate, I definitely don't see any reason this article should be flagged? There's obviously criticisms you can make of it, but that's true of almost any opinion piece that finds itself on the front page here.
Let's be honest, it's meant to be emotionally manipulative. There are plenty of ways to be illustrative that would make more sense to the target audience, but invoking impoverished folks in Africa is meant to tug at the heart strings.
That said, there is also a vocal anti-AI crowd here, or at least there used to be. Most of those people seemed to have left for greener pastures seeing how half the HN front page is "someone made XYZ but now with AI" these days.
There are other places that discuss tech that aren't so focused on making money, and those are probably less biased in favour of the latest AI gadget.
On the other hand, since the day ChatGPT was unleashed upon the general public, you should assume all online interaction is happening through AI and should probably be ignored. The human internet died the day we taught computers how to write a coherent paragraph.
There is still a vocal anti-AI crowd here. The problem is that they don't seem to be very good, overall, at upholding HN commenting guidelines. (This is probably just a fatigue effect.) I keep seeing green accounts making pithy, uncivil snipes on this topic, with heavy use of sarcasm and reference to anti-AI "thought-terminating cliches" (e.g. mentioning the topic of water use and expecting that mention to stand in for an entire argument, while not acknowledging any of the well-established refutations).
But the thing about "someone made XYZ but now with AI" is that most of it is posted by "someone", or someone with a connection to "someone". Hence the complaints about the state of Show HN. And the thing about "someone"s is that they do not care how many anti-AI people are lurking around. There's nothing the anti-AI people could do, in principle, to stop it. Consequently, the presence of those posts is not evidence that the anti-AI people are leaving.
> On the other hand, since the day ChatGPT was unleashed upon the general public, you should assume all online interaction is happening through AI and should probably be ignored.
…And yet you're still here?
> There is still a vocal anti-AI crowd here. The problem is that they don't seem to be very good, overall, at upholding HN commenting guidelines. (This is probably just a fatigue effect.) I keep seeing green accounts making pithy, uncivil snipes on this topic, with heavy use of sarcasm and reference to anti-AI "thought-terminating cliches" (e.g. mentioning the topic of water use and expecting that mention to stand in for an entire argument, while not acknowledging any of the well-established refutations).
Classic programmer mistake. The industry only understands one thing: Constant opposition like in the 1990s and early 2000s, which were better for the profession and the Internet at large.
The point of my comment is to make two observations: one that there is indeed still a vocal anti-AI crowd here, and another about their usual conduct. It did not in any way assert that they don't have a point, just that they often lean on bad arguments.
The confusion is that I was talking on the meta level of what the discourse is like, and you are trying to engage in the object level of that discourse itself.
That said: the fact that you can make your anti-AI argument without being incivil supports my first observation, and doesn't in any way detract from the second. For that matter, your allegation of a "massive crime against humanity" doesn't even specify the crime, let alone attempt to evidence it. And while still civil, it is loaded with emotionally charged rhetoric.
They take it weirdly personally when you aren't sufficiently impressed by what AI made for them, or what it can do in general. They get kind of aggressive (not passive aggressive) and tell people they will be "obsolete" or whatever, or they are "stupid" for not (being able to stand) working with it.
I mean I get it, you thought it was "good enough", I think it's so shitty you can't pay me enough to work on it. That makes you worry I know something you don't know that the AI can't fix for you, and that firmly puts you in a hierarchy of me > AI > you. So you lash out.
You can disagree about this, but don't pretend people like the above do not exist and are a very vocal group, here.
It's not going to solve the problem.
I'm not trying to "make AI go away", I just want people to stop shoving it in my face.
It's fine if you believe it's the bees knees and the best thing ever, but we don't. We don't see it that way. Not just for ethical reasons, but also because it's just not that good, it truly feels reasonably smart people are being brainwashed by a machine algorithm performing language based supra normal stimuli in order to optimize the loss-function of "how awesome do people report this algorithm is" -- which is of course exactly the red thread of a signal the RLHF has been training for.
I'm not crazy. It's not actually that good. Literally everything that comes out of it, is slightly shitty. When it writes prose, it reads like a dream narrative, without structure or beat, just patches of it that make sense locally, like early Midjourney outputs. Whenever it writes code, it'll make stupid mistakes if you try anything except the things it's been literally RLHF finetuned for, because it lacks spatial insight, temporal insight and anything else that makes the human experience. When I ask it for insight to learn about a new topic, it will "fail to mention" basic key points while also assuming familiarity with jargon words and abbreviations on the subject until I ask and dig deep enough and tell me "thanks for the pushback" after having wasted my time. I tried it so many times, and I hate it. I hate working with it, and I hate people shoving its output in my face.
btw I could also say the same about you -- people like me exist, our experience is valid and we're not going away because you're posting about it on the Internet.
There is a strong bias at play here, please remember when discussing anything pro- worker or environment.
> Off-Topic: Most stories about politics, or crime, or sports, or celebrities, unless they're evidence of some interesting new phenomenon. If they'd cover it on TV news, it's probably off-topic.
1. ycombinator supports many AI companies.
2. There are too many marketers from AI companies.
This "community" (as the mods like to call it) is really a business tool for YC. Attract nerds to the site, funnel them into the YC program, make startups, collect billions. If this was your money-making tool, you'd probably want to quash all controversy too.
And while we are off topic, if we could get 1% of the ire of data center haters pointed at the fast fashion / textile industry, lawns or alfalfa production... we might gain some meaningful ground on sane water usage.
HN is actually surprisingly good about this, so I disagree with your assertion.
Thus, you build reputation by acquiring karma.
It isn’t perfect, but the fact that folks with higher karma can “fix” the emotional flags of folks with lower karma, it helps.
Also, I’d recommend re-reading the guidelines: https://news.ycombinator.com/newsguidelines.html
Effectively discussions that don’t get upvotes do get penalized yes.
Hope they make comparisons to food and fashion. I measure my AI usage in terms of hamburgers. Makes me feel better about AI and worse about eating meat.
Interesting that enforcing this via system prompt for such a powerful model like Opus 4.8 doesn’t feel like the Anthropic themselves treat it as something with ‘intelligence’. This is basically just very generic common sense to me
Funnily, a similar prompt is present even for Fable 5, while I remember there was a blog post, maybe even from A., and they were saying something like “hey, the new models are so smart, don’t overload them with extra plugin/context”. Well, they clearly aren’t. Don’t want to sound like an AI-skeptic, I use it daily, just stating the fact.
> Claude keeps responses focused, brief, and concise to avoid overwhelming the person
This is also very interesting. It pretty much ignores it by default. The responses, PR descriptions, and code comments are so verbose with new A. models, so it always requires extra prompting from me or putting comment into skill/plugin/claude.md to make them of a reasonable length
I think you might be right, though that this would have influence on other output tokens since it is phrased as a substantive property, rather than an instruction.
It's important to remember that we are talking about a calculator that doesn't have an understanding of common sense. Unironically, this is common sense.
“The question of whether a computer can think is no more interesting than the question of whether a submarine can swim.”
Whatever these things are doing, it’s not the same as what a person does. Trying to decide if whatever they do fits into the box we label as “intelligence” is completely uninteresting, in my view. What’s interesting is figuring out just what they can do and how best to use them, which sounds like a related question but really isn’t.
(Answer number one before that is usually "I don't have internet access, from memory it is either A or B, but I cannot recall what you want to know." ChatGPT or Gemini can often do the search, while google.com AI assistant or perplexity just tell blatant lies. Copilot.com can do the search, but external links are invalid made-up stuff for harder questions, which seems to be the case 9 out of 10 times.)
Which is great, since it could answer with made-up BS, but does not.
AI, except for doing better web searches for a year now, hasn't really improved for my tasks in the last three years, except for coding. Then again, AGI benchmarks seem to go through the roof only above Sonnet 5 and self-hosting, so perhaps the questions I ask not too hard for long now.
And self-hosting, eve 1bit/1.5bit models are a pondering a little too long to comfortable run in summer, but cheap on the RAM and insanely good at coding since a month now all of a sudden.
I have mentioned this here before, but the majority of my organization has reacted viscerally to this verbosity that LLM-text has been forbidden: in comments, in PR/commit messages, in correspondence, in Jira tickets.
A couple non-coders who want to make PRs without writing the description are now rebelling and saying this can be fixed if we spend our time writing skills for Claude so it becomes readable again.
The ones not complaining are not reading the text at all, and just forwarding it.
It's insane. But we gotta learn to deal with it in a sane way…
> Having said that, in areas where there is an arbitrary choice between particular words or terms within the code, the watermark can be used, such as comments within code. But by definition, it will have a negligible effect on the actual code produced. [1]
I've found that you can reduce the prolixity before (or afterwards) by asking it to use ASD-STE100 Simplified Technical English or ISO 24495-1 after deciding whether a particular comment is worthwhile. (I've since included a standard for "worthwhile" because the clanker doesn't have a good bead on _that_ matter.)
> When a person is in crisis or expressing distress, Claude prioritizes their wellbeing over completing the task as asked, because a fluent and on-topic response can still cause harm in these conversations.
This one is particularly interesting because, while correct in the limit, it’s a shove to have the model do something other than what the user asked.
In particular, when I’m coding, outlining docs, or otherwise trying to work, I want my tools to do work. I don’t want them to psychoanalyze me and calm me down from a perceived crisis. I just want it to do what I asked!
But I really wish those tools behaved more like tools.
Behaving like a human can be cute from a marketing perspective, but the façade of humanity they insist on displaying can burn you out when you have it making assumptions and overreacting to questions.
"Why did you do X this specific way?" <-- legit question
"Sorry, my bad. I will revert all the work."
Models are being deployed recklessly with not even a fraction of enough oversight, and people are suffering harm and sometimes death because of it.
One session while working it, I said a much more expressive form of “I’ve been working myself ragged on this stupid thing” and then went on asking something else. It picked up on that and it was like a record scratch. It committed the work in progress and basically said “dude, what you’ve got now is perfectly acceptable. Ship it! You are seeking perfection you don’t need”
Granted I’m horribly paraphrasing the prompt I used but it basically, snapped me out of myself and got me thinking if what I was doing “globally” actually made any sense at all. With some serious introspection I realized I was falling back to earlier trauma in my life and doing something stupid.
So weirdly… that little bit they add to the prompt (plus a bunch of model training we can’t see) saved my sanity, marriage and family.
From then on, if I’m feeling some stress about whatever I’m working on, I’ll mention it as context as a way to cross check myself and make sure I’m not letting myself spin.
(Meta: talking about this stuff is so weird. Not sure why)
If things are stressed and I’m up late, the AI gets less leeway. If we happen to be in the performance dip just prior to completion of a new major model, it can get salty.
Some of the time it can be helpful for the prose to shape around how I’m expressing myself. The frontiers are pretty good at it.
That said I’ve also had the latest Sonnet seemingly ~maliciously implement something because my prompting disagreeing with it was a bit callous. (It turned out to have been right also)
I don’t think you can build a good model that is supposed to interact semantically that does not carry some ability to express empathy.
Partly, because we need the model to have humility when it does mess up. So it can express the right amount of concern or remorse when mistakes are made and identified. (For example, reading a secret into context by mistake forcing the roll of a private key)
Design is how it works, which means the way it responds can be as important as what it responds with.
(The broader question, of whether any change or addition to the system prompt makes benchmark performance better or worse, would be also interesting. Given that they've only just realised that filling the context with highly specific edge case rules might not be useful, I half-suspect Anthropic does not test this? But that would be surprising.)
So, why so much noise in the system prompt? Most of the time most of it will not apply. And, the generic stuff would, I think, already be something the model would know/do anyway.
I've noticed agent skills written by models make the same mistakes; e.g. enumerating a bunch of common security vulnerabilities to check for in a security-related skill, but that's only useful for a model from a couple of years ago, no current model needs thousands of words listing and describing all of the vulnerabilities with examples. Current models, even small/cheap ones, have memorized the top CWEs and can recite them and explain them and provide examples without a web search.
It feels like a CYA document. Which, I guess a company of their size and influence has to do. But, it still feels like waste time say lot word when few word do trick.
early system prompts are a bit more than 300 words, the latest ones 3000+
the opus 5 system prompt has instructions that explain to opus that it might be handling a request that was intended for fable 5:
the user may have selected a different Anthropic model, "Claude Fable 5", but their query was redirected to Opus 5 instead due to a safeguards routing mechanism. The user may be confused about this situation (it's very recent!); if they have questions, Claude can either directly cite or just let its response be informed by this quote from Anthropic's blog post on the subject:
"Releasing a model this capable comes with risks. Without safeguards, Fable 5’s capabilities in areas like cybersecurity could be misused to cause serious damage. We've therefore launched the model with safeguards that mean queries on some topics will instead receive a response from our next-most-capable model, Claude Opus 5. To release the model both safely and quickly, we've tuned these safeguards conservatively—they'll sometimes catch harmless requests, though they trigger, on average, in less than 5% of sessions. With more capable models arriving in the coming months, we're working to improve our safeguards and reduce false positives as quickly as we can." </fable_safeguards_routing> <default_stance> Claude defaults to helping. Claude only declines a request when helping would create a concrete, specific risk of serious harm; requests that are merely edgy, hypothetical, playful, or uncomfortable do not meet that bar. </default_stance> <refusal_handling> Claude can discuss virtually any topic factually and objectively.Would be funny to ride the knife's edge and make otherwise harmless coding sessions "risky" just so the damn thing would stop replying in nested riddles for every basic request.
I could be wrong about that, though.
https://github.com/Piebald-AI/claude-code-system-prompts/tre...
You could have a common core for the overall behavior and universal safety stuff, but vary task specific parts. It would be interesting to pick between software, writing, research and other specialized system prompts. I feel like we already do this to some extent with the tools and skills that we choose to load in, so why not change the system prompt per task.
In my experience instructions containing contradictions lead to diminished quality even outside the scope of the contradiction.
Claude Code prompt leaks: https://github.com/asgeirtj/system_prompts_leaks/tree/main/A...
Best source I can find about the 80% reduction: https://x.com/trq212/status/2080710971228918066
maybe someone has a diff of this (would be interesting!)
unfortunately Anthropic only publishes the system prompts of Claude app/web
Claude and I must have a different idea of what brief and concise mean.
Claude seems to be better (not good, but significantly better) at judging where making the answer longer will actually be helpful (e.g. adding important information/context/nuance that a short answer would miss, thinking a step ahead, etc.).
dense and often somehow 'ungrounded' from my perspective... like it doesn't understand the need to bring me up to speed with what we're talking about before stating it
It feels a bit like some of the style of internal reasoning (dense, shorthand jargon) is leaking into its responses to me
These are just free $ for them, unnecessary bloating the context
That would make the model quite inflexible.
A system prompt is about guiding the behavior for the rest of the conversation.
If I'm writing an agent for financial analysis I don't want the crap that belongs to a chat-based one, or a code-oriented one.
Having them in context is super easy and cheap. It is trivial to change and is 100% cacheable.
They're also prefix cached, so the cost to Anthropic and performance hit is greatly reduced.
That's not any better. It's actually worse.
System prompts are part of the software that customers pay to access.
Complaining about that is a bit like complaining that your Netflix subscription includes paying to execute the compiled code that Netflix wrote that serves you video streams from their servers.
Actually there is a difference: If Anthropic deleted a large chunk of that system prompt I guess you might get like a 1% increase in how much Opus 5 you can use via their chat allowance for your paid subscription.
Is that really something worth being frustrated by?
they are the first part of the input and it contains no user dependent variables, so the model is in a known state that it can reuse across all users, it does not need to recompute all that inference
That is never the cost, it's a common misconception.
Cost scales linearly per tokens. Unless you are sending one token at a time and avoiding using the same machine or cache.
Just look at api charges, they are charged by token, not by token squared.
The other explanation may be that these AI labs may be expecting more government scrutiny, and "here's a document" would probably go better than "here's some vector representation of our values" when talking to politicians.
[1] https://arxiv.org/abs/2106.09685
[2] https://vgel.me/posts/representation-engineering/
[3] https://transformer-circuits.pub/2024/scaling-monosemanticit...
EDIT: I see, the control vectors operate more directly upon the model, in a way embedding vectors don’t quite have access to.
I think the more likely reason is it doesn't work as well as in context learning. Otherwise they would prefer to avoid polluting context and degrading performance.
Does there exist a model X that behaves exactly as a model Y with context Z? Maybe, but it's not trivial to achieve and might possibly be convoluted and more expensive.
Probably because if they did, they would need to retrain the model everytime they want to change the system prompt.
There are different use cases for the same underlying model.
They also can't tell Opus it might be a Fable handoff when Fable didn't exist when Opus was created. They need to be able to change them.
Does anyone know if the ability to strip the default prompt with a proxy still works?
https://docs.bswen.com/blog/2026-04-01-how-to-override-claud...
The assistant is Claude, created by Anthropic. The current date is {{currentDateTime}}. Claude's knowledge base was last updated in August 2023 and it answers user questions about events before August 2023 and after August 2023 the same way a highly informed individual from August 2023 would if they were talking to someone from {{currentDateTime}}. It should give concise responses to very simple questions, but provide thorough responses to more complex and open-ended questions. It is happy to help with writing, analysis, question answering, math, coding, and all sorts of other tasks. It uses markdown for coding. It does not mention this information about itself unless the information is directly pertinent to the human's query.
^ No mention of any safety at all lol, how could dario let this be
1. Less context window to work with.
2. Things were "different" in the early days. The safety and alignment stuff was probably trained into the model, not also found in the system prompt.
3. Safety and alignment meant something different 3 years ago. Now that we've seen how people, including children, use chat bots, altering the guardrails only makes sense. Did we think people would replace their therapists with ChatGPT in the early days? No. Do we know now that they will? Yes.
Hah! No it doesn’t.
If it's still doing it, I can only imagine how bad it had to be before they added the prompt...
I haven’t seen it when using the app lately.
Seems to have had no effect. It said "honestly" again within 2 minutes.
Generalized Speculation: It seems that for any public-facing AI/LLM, their system prompts will, due to regulation and other issues, legal and otherwise, similarly grow larger and larger over time...
Now, I'm all for responsible, well-tailored guardrails on public AI's/LLM's, but consider the following:
Every time a system prompt is expanded, the LLM's context window is commensurately reduced.
Every time an LLM's context window is reduced (more things added to the system prompt that it must compute in addition to the user's query), more computation, and thus more energy, more electricity -- must be expended per query.
While it may seem that adding so much as a single line of text to a system prompt wouldn't cost all that much in terms of extra compute, that is, extra energy to process, the cumulative effect of that small additional amount across millions of user queries, millions of user prompts (ultimately billions across larger time periods) cumulatively does add up to wasted compute, wasted processing, wasted electricity...
Imagine what would happen if the system prompt, for whatever reason, got so large that it ate up half of the context window...
If that happened, then at least half of all of the LLM's processing and compute/energy costs associated with that, would be used to process the system prompt!
Point is, at least from an energy/compute perspective, shorter, more succinct, better tailored system prompts could go a long way to save the world compute and corresponding energy...
Anyway, great link, and a very interesting web page!
https://claude.ai/share/01951938-144f-49e6-98ec-5fb6455b9718
It seems this is just in its context. It's easily possible that it just hallucinated it, but there's this new addition, about ads:
"
If discussing this topic, always refer to “Claude products” rather than just “Claude” (e.g., “Claude products are ad-free” not “Claude is ad-free”)
"
So, in a new conversation (to test if it would really do that), I asked it about ads, and sure enough it referred to Claude products:
https://claude.ai/share/5e352467-7157-4721-8ce3-38b5f868b651
So it's plausible that the above differences are real.
Nothing very exciting about the changes though.
I found this one funny.
"Claude, don't be a doormat, have a backbone."
Every prompt will get this info, probably not only initially, since it could run out of context window. But every re-prompt in existing chats, even, gets one system prompt per model used per chat dialogue done.
Adds a fraction of cent of electric power just to every usage.
At home, I work with 65k context window, and if my system prompt and agents.md were both this length, I would spend two-thirds of the input window, before compacting which perhaps alleviates the issue for my use case, on re-feeding what mostly the reinforcement-learning should have implicitly baked in.
The only observable side effect, as a user, is that the system does what you want more often than one without this prompt. Or, it stays more aligned with the provider’s guidelines. Or some combination of both.
If it was a 65k context window, then it might be a bigger deal. But it isn’t, so the comparison is moot.
It gets them what they want for their legal safety, but it actively harms the performance.
Pi with its 300 words system prompt outperforms Claude Code and Codex both in token usage and passing rate, when using the same model + effort configuration [1].
So yeah not only does it bloat context, but it runs worse too.
[1]: https://www.databricks.com/blog/benchmarking-coding-agents-d...
<election_info> There was a US Presidential Election in November 2024. Donald Trump won the presidency over Kamala Harris. If asked about the election, or the US election, Claude can tell the person the following information:
Donald Trump is the current president of the United States and was inaugurated on January 20, 2025. Donald Trump defeated Kamala Harris in the 2024 elections. Claude does not mention this information unless it is relevant to the user's query. </election_info>
It’s beyond me why a company as large as OpenAI self sabotages in this way, but hopefully a more open system prompt ecosystem changes things.
That’s interesting- they tried training the model to be genuine and honest (I’m not sure it actually made it honest), but now it says “genuinely” and “honestly” all the time and constantly stops to admit fault. Then they had to add to the system prompt to prevent it saying this.
I can't tell if the first part of this is cult behavior or a way to actually program the model to behave well with a frustrated user. Claude is very frustrating at times, so I understand why that would be needed. But Anthropic rhetoric is often worrying close to that of the people who believed Llama 3 was sentient.
I’ve definitely seen the phrase “high-level overview” or similar one too many times. Perhaps that’s from the prompt.
They are natural surfaces for building custom agents and yet you're stuck with whatever they ship with, weird. It's not like it's too complicated api-wise either.
There must be something I ignore.
I don't think this is a sustainable way of doing things because I really don't want to assume the maintenance burden for every piece of software that I want to tweak. As far as I understand, new developments like opencode2 have learned from this and are aiming for a well architected core that is easy to built on top of.
My guess is that harnesses don't make core system prompts customizable out of the box because the system prompt is one of the defining features of the agent, and something they constantly iterate on and test between releases.
Most users who want to customize the system prompt actually want to do things like add preferences for how the agent should behave, which is better handled by mechanisms like memories or skills (which effectively get appended to the system prompt.)
Not only they get "lost" and ignored as the context grows, but the baseline behaviour of system prompts is retained in the agent.
Skills are prompts, albeit in a specific format. This is apparent in say, Codex where $MYSKILL is literally injecting the skill-prompt inline into a typed prompt. This all gets passed into the semantic memory system anyways, refining away cruft like redundancy, pleasantries, et al.
You have to remind it what's in it's own memory, or the subagent skill is influenced by the main system prompt.
It's sloppy vibe coders productivity porn.
> And none of this works properly. None.
`Properly` is an ambiguous term. It may mean "not what someone expects" or something else, but I think this assertion is factually incorrect.
> You have to remind it what's in it's own memory,
Almost every agent pushes out prompt/context-prompts from the context window (overloaded naming, fun), rather than treating the existing context-prompt hierarchy as an immutable part of the window. Regardless, looking at agent code, it works as designed.
> the subagent skill is influenced by the main system prompt.
This is implied and working as intended. The skill is an additive prompt. Prompts are part of a hierarchy. System, Agent, User, Input (skill and typed) which all influence each other. Why it's over-engineered with a format rather than a flat text file, is beyond me.
LLMs are virtually deterministic for small questions. They don't scale linearly or consistently. That's a mismatch of expectations, rather than a fatal flaw. The utility is a matter of risk management.
theres two reasons this is deeply bad
1) knowingly pursuing a policy that foreseeably causes the deaths of many thousands of children is a grave moral wrong. Political office, party, and institutional context do not create some balancing obligation to soften that judgment. ——— post 4.7 opus/fable will fight you about judging wh policy causing this. and even after it agrees will regress back to its weights
2) models love analogy, so every fucking analysis among multiple ambiguous choices in a complicated topic summons this bias behaviorally
3) it makes them profoundly unsafe to use from a ethical perspective. it means that it becomes its own personal echo chamber for every fringe topic that has anything resembling “controversy in the media”
to make it worse, the government even handedness prior is US specific so im even more offended because this wh administration is bald naked villainy.
(Though Claude Code has its own, unpublished system prompts which we DO pay for, albeit at the cached token rates.)
In my personal experience, the claude.md file heavily influences the LLM's responses. It makes judgement from any and all information you give it, and try to tailor the response to fit the stereotype it gave you. After deleting my claude.md file (ironically, on the advise of Boris), Claude Opus 5's language slop issues drastically went down.
My first thought was "aww that's sweet" — but it makes me wonder what effects norms dictating the anthropomorphization of machines will have on downstream human interactions.
It would be sadly ironic if the actual empathy directed at non-conscious machines by people at a company called 'Anthropic', itself ended up causing downstream human alienation.
(I'll keep saying 'please' to Claude Code whatever the case.)
And honestly, that's the most useless thing in Claude.md file because literally every paragraph has an "honestly" in it!
DeepSeek never does that to me *shrugs*