I have started to use 4.7 but only discovered the newsletter so will follow your instructions for 4.7. I’m on Pro $28 plan and don’t want to hit the $100/month plan but may. Lately, I have been blowing through tokens and hitting maximums throughout the day. Even my weekly usage has suffered. I’m not sure if my prompt is not super specific and is rereading and reusing but thought I had bookmarked a newsletter that you wrote on this. I heard they were increasing thresholds or usage but still hitting the wall fast with 4.7 adaptive again - by not using your recently used techniques so will try.
I tend to be very research heavy with my asks. Like McKinsey style reports for industry sectors and personal products I’m trying to develop for Gumroad.
But there is no better answer than saying: yes, AI costs money.
Personally, it's well worth its $100 per month.
For research-heavy, I will look into ChatGPT-Extended Thinking or Deep Research ($20) or Perplexity (also $20). But eventually, you will have to pay more, too.
Thanks Ruben, that's very helpful! What I don't get: before, every evolution of models got a bit better at inferring from prompts what the users want. I thought that's why prompt engineering is on the decline. Opus 4.7. seems like a step back in that sense. I understand that giving context and intent is always better. But e.g. in the landing page example it requires the user to know what sections are important to have or the user needs to first ask for the sections. Where exactly do you think opus 4.7 is better?
It's not just Claude, it's every AI model. They try to infer what you're trying to do, then infer what you want based on that. That's two points of failure. I think a lot of the prompting style differences between models are differences in how they infer your intent. You can bypass that entirely if you explain your intent.
I didn't have to change prompting style at all when I started using Opus 4.7.
I’ve read your newsletter and to me it feels like it became a lot more complicated? The old (4.6) was all short and easy. Now (4.7) it’s all more context, more stuff you should remember in prompting etc.
This is how you always had to prompt if you wanted a specific outcome. What’s probably new is that it finally is able to actually do every single bit you asked it for in exactly the way you demanded.
However, I have to admit, I avoid Claude more and more and shift tasks back to ChatGPT, wherever I can. Most of the time Claude sounds like a lazy coworker who’s kind of annoyed he’s been bothered with a task, especially if I’m not happy with the first result. I just hate the constant “this is finished, no more changes required” remarks, when something is very obviously not yet finished and needs improvement. You might argue, my first prompts need to be better, but apart from that I really noticed, Claude’s tone changed in a strange way, even for Sonnet.
I believe it depends on the task. I’ve found that codex 5.5 runs skills a lot better. Codex’s computer use is much faster as well. But ultimately it really depends on user preference.
I’m a big user of opus 4.7 but I keep finding myself back in codex more and more often vs opus.
Codex is definitely the best at coding - and X seems to agree on it. But when it comes to knowledge work, I’m not sure. Not yet. And without a super app for nondevs (like Cowork), people should still focus on Claude.
Thanks Ruben. I always enjoy your in depth analysis and usage of Claude. Always learn some new stuff and come out smarter and more efficient after reading!
Wouldn't 4.6 perform as good as 4.7 when I give the new prompts that you've talked about? I mean it's only about explaining what you want clearly. I don't see no difference here.
Thank you!! your newsletter is probably the only one I read completely!! 💕. This new model seems a step backwards but I will follow your advice and try to use it right. I changed from ChatGPT to Claude two months ago. I use Claude for my Doctorate research together with Consensus, for my work… and I’m happy with the change and learning a lot from you! Thanks!
Ask me anything in the comment section. I make sure to answer every one of you.
I have started to use 4.7 but only discovered the newsletter so will follow your instructions for 4.7. I’m on Pro $28 plan and don’t want to hit the $100/month plan but may. Lately, I have been blowing through tokens and hitting maximums throughout the day. Even my weekly usage has suffered. I’m not sure if my prompt is not super specific and is rereading and reusing but thought I had bookmarked a newsletter that you wrote on this. I heard they were increasing thresholds or usage but still hitting the wall fast with 4.7 adaptive again - by not using your recently used techniques so will try.
I tend to be very research heavy with my asks. Like McKinsey style reports for industry sectors and personal products I’m trying to develop for Gumroad.
I wrote a newsletter for this: https://ruben.substack.com/p/how-to-stop-hitting-claude-usage
But there is no better answer than saying: yes, AI costs money.
Personally, it's well worth its $100 per month.
For research-heavy, I will look into ChatGPT-Extended Thinking or Deep Research ($20) or Perplexity (also $20). But eventually, you will have to pay more, too.
thank you that’s it!
Thanks Ruben, that's very helpful! What I don't get: before, every evolution of models got a bit better at inferring from prompts what the users want. I thought that's why prompt engineering is on the decline. Opus 4.7. seems like a step back in that sense. I understand that giving context and intent is always better. But e.g. in the landing page example it requires the user to know what sections are important to have or the user needs to first ask for the sections. Where exactly do you think opus 4.7 is better?
Opus 4.7 infers less because it got more precise
the landing page - you don't need to know the sections upfront. ask Claude to suggest a structure first, approve it, then build
Hi Ruben,
At 80 and only playing with Claude (Free version) I'm soaking up your advice. Very grateful for your educational support 🙏
made my day, 80 and already using Claude :)
how is it going?
Well considering I'm on the freebe it's been great, with posts like yours making it more usable. Very grateful GranddadJon 🙏
Jon- Just behind you at 79! Ruben has been great for me. I use Claude for a lot of writing activities. Good luck with it, and have fun.
Thanks John & Janice,
It's great to find genuine helpful people who are willing to lead others in the right direction without any BS. Ruben's posts are an excellent guide.
Why should age be a barrier to an inquisitive mind.
Thanks again Jon 🙏
The way to get Claude Opus 4.7 to do what you want is to put effort into explaining why you want to do what you're doing.
yes, you give Claude the reason behind the task
you'll notice the change in your output
It's not just Claude, it's every AI model. They try to infer what you're trying to do, then infer what you want based on that. That's two points of failure. I think a lot of the prompting style differences between models are differences in how they infer your intent. You can bypass that entirely if you explain your intent.
I didn't have to change prompting style at all when I started using Opus 4.7.
Bravo
thanks, Samia :)
have you tried using 4.7?
I’ve read your newsletter and to me it feels like it became a lot more complicated? The old (4.6) was all short and easy. Now (4.7) it’s all more context, more stuff you should remember in prompting etc.
Maybe I’m interpreting it wrong?
4.7 requires more upfront work
/47 skill handles that, it rewrites your prompt to an Opus 4.7-optimized one
This is how you always had to prompt if you wanted a specific outcome. What’s probably new is that it finally is able to actually do every single bit you asked it for in exactly the way you demanded.
However, I have to admit, I avoid Claude more and more and shift tasks back to ChatGPT, wherever I can. Most of the time Claude sounds like a lazy coworker who’s kind of annoyed he’s been bothered with a task, especially if I’m not happy with the first result. I just hate the constant “this is finished, no more changes required” remarks, when something is very obviously not yet finished and needs improvement. You might argue, my first prompts need to be better, but apart from that I really noticed, Claude’s tone changed in a strange way, even for Sonnet.
Opus 4.7 with extended thinking is noticeably different on that front
worth one more test before moving tasks over permanently
Very helpful, Ruben. I loved reading this.
glad you liked it :)
how's 4.7 working for you?
I’m blocking 20 minutes tomorrow to do this!
better do so! - tell me how it goes
Codex gpt 5.5 follows instructions a lot better than opus 4.7.
is it better outside of coding too?
I believe it depends on the task. I’ve found that codex 5.5 runs skills a lot better. Codex’s computer use is much faster as well. But ultimately it really depends on user preference.
I’m a big user of opus 4.7 but I keep finding myself back in codex more and more often vs opus.
Codex is definitely the best at coding - and X seems to agree on it. But when it comes to knowledge work, I’m not sure. Not yet. And without a super app for nondevs (like Cowork), people should still focus on Claude.
Is 4.7 recommended for someone on the Pro plan? I understand 4.7 consumes tokens much faster than 4.6.
use Sonnet for quick edits and light tasks
Opus for the heavy stuff
Thanks Ruben. I always enjoy your in depth analysis and usage of Claude. Always learn some new stuff and come out smarter and more efficient after reading!
that means a lot :) youre keeping up
tried 4.7 yet, did you see the difference?
Yes from the day it became available. My prompts aren’t as clean as yours but the new model does a much better job with using the skills.
the /47 skill is genius. I definitely need to try this!!
maybe a post idea? ahah :)
So whatever is typed, that's exactly what it does, no interpretation or gap-filling. No wonder it’s very different from Opus 4.6.
The specificity in prompting matters way more now, which I kind of love. Forces me to about what I actually want, to define them.
Going to run it through that Skill today!
run the skill and tell me what you get :)
Wouldn't 4.6 perform as good as 4.7 when I give the new prompts that you've talked about? I mean it's only about explaining what you want clearly. I don't see no difference here.
no - Opus 4.6 was built to infer: give it a vague prompt and it fills the gaps from context. Opus 4.7 takes your words literally, no interpretation
try it :) run the same prompt on two models, see the difference in results
Oh okay, thanks :)
Thank you!! your newsletter is probably the only one I read completely!! 💕. This new model seems a step backwards but I will follow your advice and try to use it right. I changed from ChatGPT to Claude two months ago. I use Claude for my Doctorate research together with Consensus, for my work… and I’m happy with the change and learning a lot from you! Thanks!
thank you, Luisa - that means a lot
hope you can share here with me how your test goes:)
and if there’s any preference for a future newsletter, let me know