Karo, I swear you are a linguistic Surgeon, it's amazing that you're able to break all of this down from that tweet and to be able to share it with all of us in public as well. Bookmarking this for later, as I've just entered the world of vibe coding, inspired mainly by your wonderful posts. Thanks again for the reminder of what I can hopefully achieve in the not-too-distant future.
For sure, I've got a Substack planned about this in the next couple of weeks. I can even send you what I've built so far if you want. Obviously @Karen Spinner has been holding me hand!
Great summary! He shows that you don’t need to overcomplicate things to make the best use of the tool, sometimes less is more. I also agree with how he thinks about agents: the smaller the task, the better, since it’s easier to reuse and debug if needed. I believe tiny tasks in the form of agents are the way to go.
The shared claude.md file is so interesting! We have been doing something similar with a shared agents in a marketplace plugin but haven’t thought about the approach as claude.md feels so personal!
Thanks for breaking it down in such clear way, Karo.
There's the counter argument some people have that bcuz he is from Anthropic, not only he has to say CC is the best tool out there but he has almost unlimited usage so he can pull off things most of us mortals can't.
I appreciate that he's pretty transparent in his way of using it with the little details (even if some of us can't replicate that cuz we would max out on tokens)
That’s a very valid point. I was thinking of including it as well, but wanted to stay true to what he actually shared. He has unlimited credits, so his exact setup isn’t accessible 1:1 to most people. Tokens aside, he shared some patterns we can copy & test.
I've been tinkering with Claude Code for months, but I only scratched the surface of what it can do. Therefore, I can't wait to put some of these into practice!
Interesting to hear an AI guru say this. I know they have disclaimers, but when are they going to tell you -- do not trust one single thing this machine says? Go and fact check every single little tiny item, because these bots fabricate and lie about everything? Especially do not trust the bot when it says "Trust me, I've checked this"?
When I asked DeepSeek whether one extended narrative in a research exercise was really true, it assured me it was and said "I have fallen on my sword for less". The whole thing turned out to be a huge hallucination, a vast pack of lies that it lied about and then lied further about its lies. Who needs a tool that deliberately lies to you? And then lies to cover up its lies?
It's the best use case when the creator of Claude Code is telling you how he uses it. Since he knows better than anyone how to use the tool, he can lay everything out for you. It's essentially a roadmap for how to use Claude Code to its full potential. Amazing post!
This is so freaking useful. I had no idea plan mode was a thing, and I hadn't thought of using CLAUDE.md as a knowledge base for bugs. Brilliant stuff.
Like I said, a lot of it went over my head, but I keep returning to it because I know there's some really useful information about how those of us from the analog age can access and make use of AI in an authentic way.
Thanks! Gave me some ideas to try. I’m starting to feel confident enough in Opus 4.5’s code that I want to start spinning up multiple agents like that.
It made me think about how I use LLM's in general. And in particular when should we have steering documents like Claude.MD at the team level, project level, product level, or organization level.
Karo, I swear you are a linguistic Surgeon, it's amazing that you're able to break all of this down from that tweet and to be able to share it with all of us in public as well. Bookmarking this for later, as I've just entered the world of vibe coding, inspired mainly by your wonderful posts. Thanks again for the reminder of what I can hopefully achieve in the not-too-distant future.
How great that you're starting to vibecode! Are you going to use Claude or Replit or Lovable?
Gemini! 🤣 then onto lovable I think…
Great! Will you share your process?
For sure, I've got a Substack planned about this in the next couple of weeks. I can even send you what I've built so far if you want. Obviously @Karen Spinner has been holding me hand!
That's fantastic! You're in great hands with Karen.
You built something functional and got it running on your website. That’s a win! 🎉
Thanks so much for all your help, Karen. It's fair to say I'm hooked and now need to start building things a little bit more complicated.
Great summary! He shows that you don’t need to overcomplicate things to make the best use of the tool, sometimes less is more. I also agree with how he thinks about agents: the smaller the task, the better, since it’s easier to reuse and debug if needed. I believe tiny tasks in the form of agents are the way to go.
Yes, he keeps the setup rather simple. Even calls it ''vanilla''. But then he spins 10+ agents simultaneously 😂
The shared claude.md file is so interesting! We have been doing something similar with a shared agents in a marketplace plugin but haven’t thought about the approach as claude.md feels so personal!
It does, doesn't it? No one should ever see mine 😂
This is super valuable. Thanks for sharing ❤️
My absolute pleasure, thank you for reading Nicola 🤗
Thanks for breaking it down in such clear way, Karo.
There's the counter argument some people have that bcuz he is from Anthropic, not only he has to say CC is the best tool out there but he has almost unlimited usage so he can pull off things most of us mortals can't.
I appreciate that he's pretty transparent in his way of using it with the little details (even if some of us can't replicate that cuz we would max out on tokens)
That’s a very valid point. I was thinking of including it as well, but wanted to stay true to what he actually shared. He has unlimited credits, so his exact setup isn’t accessible 1:1 to most people. Tokens aside, he shared some patterns we can copy & test.
Exactly. He at least shared his whole workflow without gatekeeping any "secret sauce".
Yes, super cool.
Gold
Thank you for reading David! 🤗
I've been tinkering with Claude Code for months, but I only scratched the surface of what it can do. Therefore, I can't wait to put some of these into practice!
That’s great! Let us know how it goes!
"You don’t trust; you instrument."
Interesting to hear an AI guru say this. I know they have disclaimers, but when are they going to tell you -- do not trust one single thing this machine says? Go and fact check every single little tiny item, because these bots fabricate and lie about everything? Especially do not trust the bot when it says "Trust me, I've checked this"?
When I asked DeepSeek whether one extended narrative in a research exercise was really true, it assured me it was and said "I have fallen on my sword for less". The whole thing turned out to be a huge hallucination, a vast pack of lies that it lied about and then lied further about its lies. Who needs a tool that deliberately lies to you? And then lies to cover up its lies?
Yes, fact checking should be normalized, for everything that we do.
Another amazing article! Gonna dive deep into Claude Code soon, and I for sure will have this article opened for reference.
That's fantastic, thank you James!
It’s a great summary! Thanks for sharing Karo! Big fan!
My pleasure, thank you for reading Adam 🤗
It's the best use case when the creator of Claude Code is telling you how he uses it. Since he knows better than anyone how to use the tool, he can lay everything out for you. It's essentially a roadmap for how to use Claude Code to its full potential. Amazing post!
Thank you for reading Ilia and I agree, I'd love to see builders from other companies do that too.
This is so freaking useful. I had no idea plan mode was a thing, and I hadn't thought of using CLAUDE.md as a knowledge base for bugs. Brilliant stuff.
Thank you for reading Privacat!!!
Wow - supercool. Thanks for the summary. I feel like I only got 30% of this - but even that was useful. Dedicated to return for further study.
That's wonderful, thank you for reading Marko!
Like I said, a lot of it went over my head, but I keep returning to it because I know there's some really useful information about how those of us from the analog age can access and make use of AI in an authentic way.
Thanks! Gave me some ideas to try. I’m starting to feel confident enough in Opus 4.5’s code that I want to start spinning up multiple agents like that.
Me too! I need to up my tab game 😂 I've never tried CC in 10 tabs simultaneously.
It *sounds* anxiety-inducing, but I'm wanting to try haha.
It does hehe 😂 But time savings > anxiety, at least for me.
Nice one!
Thanks Jos, glad you enjoyed it. Do you use CC in a similar way?
It made me think about how I use LLM's in general. And in particular when should we have steering documents like Claude.MD at the team level, project level, product level, or organization level.
Great lessons we can learn a lot from. The right strategy is everything
These really are great lessons, I learned a lot from him 🤗