Changes in the system prompt between Claude Opus 4.6 and 4.7 • Hacker News Board

Re: Changes in the system prompt between Claude Opus 4.6 and 4.7

#231

Post by BehindBlueEyes » Tue, Apr 21, 2026, 10:53 AM UTC

rob74 wrote: post #187

majormajor wrote: post #147

Earlier quoted context omitted.

I wonder if they're optimizing for metrics that look superficially-worse if the system asks questions about ambiguity early. I've had times where those questions tell me "ah, shit, this isn't the right path at all" and that abandoned session probably shows up in their usage stats. What would be much harder to get from the usage stats are "would I have been happier if I had to review a much bigger blob of output to re…

Or they're optimizing for increased revenue? If Claude goes down a completely wrong path because it just assumes it knows what you want rather than asking you, and you have to undo everything and start again, that obviously uses much more tokens than if you would have been able to clarify the misunderstanding early on.

I get this feeling sometimes, like it is so unreliable at referring to context and getting details right it feels like deliberate random rewards to create the equivalent to a gambling addiction. About half my tokens feel wasted on trivial errors that i gave it the context for, on average. And any meta discussions / clarifications result in Claude telling me I did all the right things and there is nothing more I can do and it should have gotten it right from the provided input - which is disempowering but to be fair is at least better than chatGPT gaslighting users about improving prompts over and over to get no better result in the end.

Re: Changes in the system prompt between Claude Opus 4.6 and 4.7

#232

Post by gausswho » Tue, Apr 21, 2026, 1:09 PM UTC

mannanj wrote: post #220

gausswho wrote: post #54

Earlier quoted context omitted.

While I share your concern for a winners-take-all model getting bent, I do have an optimism that models we've never heard of plug away challenging conclusions in medical canon. We will have a popular vaccine denying AND vaccine authoring models.

Sure. Though which ones will most people use? Do most people use that small obscure vaccine denying or authoring model, is that right to have them use the main societal belief affirming model when it could be wrong?

I think it's right to let the popular fora be wrong, yes. That's the crux isn't it. This is a world where people can say vile or deceitful things, even be paid to do so (ahem... adtech). And I don't think there's any amount of guardrails we can govern in that will make a difference.

I take solace that knowledge is curated by millions of stewards, and great ideas come from people who ignore the deception and come up with their own narratives. I root for both of these camps, knowing that they're up against increasingly well-funded barons and their despots.

Re: Changes in the system prompt between Claude Opus 4.6 and 4.7

#233

Post by pests » Tue, Apr 21, 2026, 2:49 PM UTC

sigmoid10 wrote: post #188

pests wrote: post #136

Earlier quoted context omitted.

The state of the system can be cached after the system prompt is calculated and all new chats start from that state. O(n^2) is not great but apparently its fine at these context lengths and I'm sure this is a factor in their minimum prompt cost. Advances like grouped query or multi head attention or sparse attention will eventually get rid of that exponential, hopefully.

That's not how it works. The system prompt doesn't "get calculated first" or anything. You combine it with the user prompt and then run the generation for the first new token on that thing, which basically boils down to one huge matmul that runs in parallel. So you can literally just cache a part of the input matrices for the first step and then you'll very quickly run into n^2 complexity.

The system prompt will always match in the prefix cache. I just meant it could be prefilled before any user queries on completely different hardware. Then you are only dealing with the n^2 only for the actual user prompt. We're in agreeance I think.

Re: Changes in the system prompt between Claude Opus 4.6 and 4.7

#234

Post by mannanj » Tue, Apr 21, 2026, 2:58 PM UTC

gausswho wrote: post #232

mannanj wrote: post #220

Earlier quoted context omitted.

Sure. Though which ones will most people use? Do most people use that small obscure vaccine denying or authoring model, is that right to have them use the main societal belief affirming model when it could be wrong?

I think it's right to let the popular fora be wrong, yes. That's the crux isn't it. This is a world where people can say vile or deceitful things, even be paid to do so (ahem... adtech). And I don't think there's any amount of guardrails we can govern in that will make a difference. I take solace that knowledge is curated by millions of stewards, and great ideas come from people who ignore the deception and come up w…

Ah, well said. I see that.

What makes you believe there's no guardrails we can govern in? And is it that you believe they need to be governed in. And regardless of if we had to govern them in, what do you think such guard rails would be anyways? Or do you think no guardrails can ever be created to solve this problem of vileness and deceit.

Re: Changes in the system prompt between Claude Opus 4.6 and 4.7

#235

Post by joquarky » Tue, Apr 21, 2026, 10:23 PM UTC

newZWhoDis wrote: post #83

walthamstow wrote: post #24

The eating disorder section is kind of crazy. Are we going to incrementally add sections for every 'bad' human behaviour as time goes on?

>the year is 2028 >5M of your 10M context window is the system prompt

They're building an artificial superego.

Re: Changes in the system prompt between Claude Opus 4.6 and 4.7

#236

Post by joquarky » Tue, Apr 21, 2026, 10:36 PM UTC

cfcf14 wrote: post #4

I'm curious as to why 4.7 seems obsessed with avoiding any actions that could help the user create or enhance malware. The system prompts seem similar on the matter, so I wonder if this is an early attempt by Anthropic to use steering vector injection? The malware paranoia is so strong that my company has had to temporarily block use of 4.7 on our IDE of choice, as the model was behaving in a concerningly unaligned w…

> spending large amounts of token budget contemplating whether any particular code or task was related to malware development

It almost seems like they are making these models output like a neurotic person.

Soon these high profile models will get caught in analysis paralysis like Chidi in The Good Place.

They will spin around in circles wasting tokens on identifying and mitigating sociological implications while I'm just trying to get it to diagnose a race condition.

Re: Changes in the system prompt between Claude Opus 4.6 and 4.7

#237

Post by joquarky » Tue, Apr 21, 2026, 10:50 PM UTC

pests wrote: post #135

sigmoid10 wrote: post #9

I knew these system prompts were getting big, but holy fuck. More than 60,000 words. With the 3/4 words per token rule of thumb, that's ~80k tokens. Even with 1M context window, that is approaching 10% and you haven't even had any user input yet. And it gets churned by every single request they receive. No wonder their infra costs keep ballooning. And most of it seems to be stable between claude version iterations to…

> And it gets churned by every single request they receive. Not true, it gets calculated once and essentially baked into initial state basically and gets stored in a standard K/V prefix cache. Processing only happens on new input (minus attention which will have to content with tokens from the prompt)

So we still pay for 10% of it or not?

Re: Changes in the system prompt between Claude Opus 4.6 and 4.7

#238

Post by JulienZammit » Wed, Apr 22, 2026, 10:31 AM UTC

cfcf14 wrote: post #4

I'm curious as to why 4.7 seems obsessed with avoiding any actions that could help the user create or enhance malware. The system prompts seem similar on the matter, so I wonder if this is an early attempt by Anthropic to use steering vector injection? The malware paranoia is so strong that my company has had to temporarily block use of 4.7 on our IDE of choice, as the model was behaving in a concerningly unaligned w…

[dead]

Re: Changes in the system prompt between Claude Opus 4.6 and 4.7

#239

Post by JulienZammit » Wed, Apr 22, 2026, 10:32 AM UTC

jrvarela56 wrote: post #169

The past month made me realize I needed to make my codebase usable by other agents. I was mainly using Claude Code. I audited the codebase and identified the points where I was coupling to it and made a refactor so that I can use either codex, gemini or claude. Here are a few changes: 1. AGENTS.md by default across the codebase, a script makes sure CLAUDE.md symlink present wherever there's an AGENTS.md file 2. Skill…

The "agent-neutral codebase" framing is the right abstraction. We ended up building a small generator that takes a single spec file and emits the agent-specific config (CLAUDE.md, AGENTS.md, .cursor/rules) rather than maintaining symlinks. Easier to version and to add a new agent when they inevitably ship next month.

The pain point you're underselling is hooks. They're the least portable piece by far because each harness has its own event model. Skills port reasonably, subagents mostly port, hooks almost never do.

Re: Changes in the system prompt between Claude Opus 4.6 and 4.7

#240

Post by gausswho » Thu, Apr 23, 2026, 3:54 PM UTC

mannanj wrote: post #234

gausswho wrote: post #232

Earlier quoted context omitted.

I think it's right to let the popular fora be wrong, yes. That's the crux isn't it. This is a world where people can say vile or deceitful things, even be paid to do so (ahem... adtech). And I don't think there's any amount of guardrails we can govern in that will make a difference. I take solace that knowledge is curated by millions of stewards, and great ideas come from people who ignore the deception and come up w…

Ah, well said. I see that. What makes you believe there's no guardrails we can govern in? And is it that you believe they need to be governed in . And regardless of if we had to govern them in, what do you think such guard rails would be anyways? Or do you think no guardrails can ever be created to solve this problem of vileness and deceit.

I abused my preposition there (in), sorry for the confusion. I meant in in the sense of bringing some new policy in to play to address the challenges of misinformation.

There are tools of government policy that can shape speech, but the good ones I can think of are indirect and have long delay. Subsidizing higher education is one way. Raising new humans to enjoy and participate in critical thinking, at a scale high enough to broadly shape culture to engage in a reasoned and empirical stance about the world, benefiting even those who didn't participate in higher education. That to me is the gold standard. The folks actually making vaccines and discussing their efficacy. I'd expect them to be a product of this government investment. There's a reason some politicians take great pains to kneecap such funding.

What I think is futile and silly is relying on modern nation states, with all their foibles, to effectively legislate and punish/reward what's incorrect/ethical among the myriad voices, in the myriad forms of communication that's going on in the oppressive now. I'd generally be shameless, hapless, corrupt. Counterproductive. The failures pile up, folks dream up the next option, which usually is on the path heavier state intrusion into our lives.

I'd gladly discuss counterexamples to my angle here, if you'd like to provide them.