I just downloaded the new set of files and the blueprint global template for claude md still doesn't match what you have in this article. I'm not sure if that is intentional or just poor quality products. I also don't see any mention of the memory system in the global template, so I'm thinking it's the latter. I honestly would like a refund.
Hey, I think you should get your refund in three to five days. Just as some words of reflection:
As for the product itself, I think it is more about me going a bit too deep into other things. I also think that some statements in my past articles change over time, and also my products and packages and everything change over time. Because it has changed over time, it is very hard to have the same kind of thing that it was before. What I mean by that is that when I was writing this and I was creating the first version of this, it was totally what it was and it was 100% what I intended it to be. It was a while ago, and after that I was doing a lot of other stuff. My approach to some other things also changed, and I improved many things in the way, and some other things I had to ditch. I think that's why there is a mismatch in this. I don't think this is something bad. It is just that I have to figure out how to make it more up to date, in terms of also communicating that kind of change. Should I leave the old products intact and create new ones or new versions, or something like that? I don't know, and I admit transparently that I don't know.
I know for sure that when something is changing, this is it, and also the quality could be a problem over time when I want to add or change things too many times. Then the quality is different from the first and the last, and this is the learning that I really have. Sorry about all of the stuff, Eric. I hope that you get at least something from the product that you got, and I will promise that I will keep improving. Hopefully I will have better things in the future that maybe you will like more.
I just paid for the package but you did not update your Claude.md file according to this article. Specifically, in the article it says you end your claude.md file with this
Thanks for this comment - I know what happened - I was updating this product recently after Claude Code leak and it slipped. Sorry about that - I am just sending you an e-mail with fixed pack(and also it i fixed for the future!).
Interesting how the optimal size depends on what you're building. For a task executor — yes, shorter is better. But I use CLAUDE.md as a shared environment for long-term collaboration — projects, context, relationships between ideas. It grows because the connections grow. Same file, completely different purpose.
Yeah, I would say that this file should be treated more like instructions and the foundation of many things that you want to do in a very specific way. Of course, many people are using this as kind of an extension of a memory system, so it is loaded every time. Your agent or AI chat has a short-term memory inside this, but after a while I would rather advise not to do that. Include in the http://cloud.md file a reference to another file, http://memory.md, that has longer context and longer things when needed, of course.
“Put everything else in named reference files the agent can load when relevant”
That’s a great piece of advice! Currently, my Claude.md is long (I don’t even know how long), and I was wondering how to optimize it as it uses a lot of a context window. I will try the approach with splitting it and referencing additional files on specific topics.
I wonder how to test evals - atm I’m happy with responses and would like to ensure that after the change it works in the same way, but with smaller context - I will ask Claude to generate evals probably 😅
Thanks Pawel for sharing it, it makes so much sense and resonate how CLAUDE.md file makes all the difference.. key thing i learned is less is more.. as you saw in my article too, I try following the Progressive Disclosure for my project files
This is exactly the kind of practical insight I needed. I've been prompting on the fly and wondering why results felt inconsistent.
The "principle beats rule" insight hit me hard. I've been writing specific instructions for every edge case, which is exactly what you said doesn't work.
I just downloaded the new set of files and the blueprint global template for claude md still doesn't match what you have in this article. I'm not sure if that is intentional or just poor quality products. I also don't see any mention of the memory system in the global template, so I'm thinking it's the latter. I honestly would like a refund.
Hey, I think you should get your refund in three to five days. Just as some words of reflection:
As for the product itself, I think it is more about me going a bit too deep into other things. I also think that some statements in my past articles change over time, and also my products and packages and everything change over time. Because it has changed over time, it is very hard to have the same kind of thing that it was before. What I mean by that is that when I was writing this and I was creating the first version of this, it was totally what it was and it was 100% what I intended it to be. It was a while ago, and after that I was doing a lot of other stuff. My approach to some other things also changed, and I improved many things in the way, and some other things I had to ditch. I think that's why there is a mismatch in this. I don't think this is something bad. It is just that I have to figure out how to make it more up to date, in terms of also communicating that kind of change. Should I leave the old products intact and create new ones or new versions, or something like that? I don't know, and I admit transparently that I don't know.
I know for sure that when something is changing, this is it, and also the quality could be a problem over time when I want to add or change things too many times. Then the quality is different from the first and the last, and this is the learning that I really have. Sorry about all of the stuff, Eric. I hope that you get at least something from the product that you got, and I will promise that I will keep improving. Hopefully I will have better things in the future that maybe you will like more.
Hi, got it. Sorry about that - I will work to improve quality! In the meantime - I will refund you. Thanks for feedback!
I just paid for the package but you did not update your Claude.md file according to this article. Specifically, in the article it says you end your claude.md file with this
"## Reference (load when needed)
- Content/writing style: docs/agent/content.md
- Memory system: docs/agent/memory-system.md
- Model selection: docs/agent/model-selection.md
- WizBoard tracking: docs/agent/wizboard.md
- Self-systems: docs/agent/self-systems.md
- Infrastructure/deployment: docs/agent/infrastructure.md"
but that is not what the paid download has. Please fix this and notify me so that I can get new files.
Hey Eric!
Thanks for this comment - I know what happened - I was updating this product recently after Claude Code leak and it slipped. Sorry about that - I am just sending you an e-mail with fixed pack(and also it i fixed for the future!).
I just downloaded the new set of files and it still doesn't match what you have in this article. I'm not sure if that is intentional.
That’s really great - thanks for sharing!
Interesting how the optimal size depends on what you're building. For a task executor — yes, shorter is better. But I use CLAUDE.md as a shared environment for long-term collaboration — projects, context, relationships between ideas. It grows because the connections grow. Same file, completely different purpose.
Yeah, I would say that this file should be treated more like instructions and the foundation of many things that you want to do in a very specific way. Of course, many people are using this as kind of an extension of a memory system, so it is loaded every time. Your agent or AI chat has a short-term memory inside this, but after a while I would rather advise not to do that. Include in the http://cloud.md file a reference to another file, http://memory.md, that has longer context and longer things when needed, of course.
There are so many parallels between writing a good Claude.md file and being a good leader of a team.
Yes! Like, being clear what you want :D
The main diference → as aleader you have to also think about whole context of the person. And for AI → it’s more about being specific.
“Put everything else in named reference files the agent can load when relevant”
That’s a great piece of advice! Currently, my Claude.md is long (I don’t even know how long), and I was wondering how to optimize it as it uses a lot of a context window. I will try the approach with splitting it and referencing additional files on specific topics.
I wonder how to test evals - atm I’m happy with responses and would like to ensure that after the change it works in the same way, but with smaller context - I will ask Claude to generate evals probably 😅
I’m building an AI agent for product management (https://productpeak.substack.com/p/your-ai-has-no-memory-mine-does) and will improve the blueprint if the change makes a difference. Thanks!
I will give it a try on the branch and compare results of the old setup vs new setup- in terms of quality of responses and context window.
I will report back what I found 😅
Thanks!
Having optimised Claude md also is smart because context decay. Even with 1M context window!
this looks an awesome read Pawel! will be taking a look :)
Thanks! From time to time I think it’s good to write about basics. Sometimes I am going too deep for some people into tech :D
And this one is useful for many people - not only Claude Code/Codex users!
the struggle is real! trying to keep everything interesting and new, but then remembering you are picking up readers where even the basics are new.
your doing a great job!
Thanks Pawel for sharing it, it makes so much sense and resonate how CLAUDE.md file makes all the difference.. key thing i learned is less is more.. as you saw in my article too, I try following the Progressive Disclosure for my project files
Super! BTW. I saw a Theo YT Video about Claude MD and Agents - TL:TR → worth testing Ai with and without this file. You might be suprised!
This is exactly the kind of practical insight I needed. I've been prompting on the fly and wondering why results felt inconsistent.
The "principle beats rule" insight hit me hard. I've been writing specific instructions for every edge case, which is exactly what you said doesn't work.
Rethinking my whole setup this week. Thank you.
Thanks for mindful comment as well!
Like a sourdough starter, it requires constant feeding.
Like a bonsai tree, it requires constant pruning.