"Keep your CLAUDE.md thin. Every line you add is a line the model has to read at the start of every single session. Treat it like precious real estate. If you can say it shorter, say it shorter."
100% this. in fact, i think this will be more important during each iteration of the models. two years ago these files needed to be big, and it seems like we can rely on the models to do more of the heavy lifting and reasoning every few months.
if you haven't taken the time to thin out your claude.md and some of your skills, now is a good time to do so as we prepare for the coming drops!
Most people optimizing their agent are still in the execution layer — shaving tokens, tuning prompts, but context loading is a different problem. Someone has to decide what gets loaded: you, RAG, or the agent itself, that choice doesn't disappear just because the window is 200k.
Thanks Pawel. This described some experiments I’ve been building without framing them expressly as agents. Claude Code, Obsidian vault, instruction files with autonomy tiers, scheduled monitors, memory as markdown. I’m about to go deeper with an old Mac Mini that’s currently just sitting idle. Was thinking of trying OpenClaw on it but maybe I’ll build something from scratch for the full learning experience.
Thanks Jonatan. Sounds like you're further along than you think. Autonomy tiers, scheduled monitors, memory as markdown... that's exactly the shape. You've been building an agent without calling it one.
On the Mac Mini question, go for it. The moment I moved off my MacBook was the moment my agent stopped being a hobby project. That old Mac Mini will run a surprising amount of stuff.
OpenClaw vs building from scratch: I'd build from scratch. Not because OpenClaw is bad, but because the rebuild is where most of the learning is. Read other repos for concepts (I was inspired by a bunch of ideas that way), but owning the architecture yourself is how you understand what's yours to tune. Let me know how it goes.
There’s also a third option I’m watching. The signs from Anthropic arguably point toward a built-in agent platform, especially what was visible in the leak. Looks like scaffolding for persistent agents inside the Claude ecosystem. I’m half tempted to wait and see what ships natively before I invest a lot of time in building my own infrastructure.
But that’s a different discussion and the learning experience will probably be worth it regardless of what Anthropic ships next. Starting small like you recommend in your post.
I tinkered around with Claude projects and I was also very surprised that in the end everything is more or less text. In Claude project you can add instructions (I guess this is the role of the agent), files for context and you have a chat window. I connected Notion to Claude and it created some websites for me according to my design (which was added as a file called brand.md) and transferred a description of the results to Notion. Another experiment I did was an extensive research and then creating Notion databases for me, which worked really well.
In this harness of Claude project, there are no folders on my laptop, it's more the UI of Claude Project, but I did not test yet something that would write into my laptop's folders. So I'm not yet quite sure where in my setup would be the Claude.md.
What I also do find confusing, is the Skills section of Claude. There are many different folders and I do not really get the point of having an agent in the Skills section vs. having project instructions. My questions are related especially to non-coding tasks like researching something for me and creating Notion databases. This all worked without the Skills section in Claude. So I guess this is perhaps relevant later.
Bianca, thank you. You nailed it in your first paragraph: once you realize everything is text, the whole thing stops being mysterious.
On where CLAUDE.md fits in your setup, you're already using the equivalent. In Claude Projects, the 'Custom instructions' field IS your CLAUDE.md. Same idea, different surface. The files you add for context are your project memory. The chat window is your channel. You have the three pieces, just wrapped in a UI instead of on disk. Nothing wrong with that, especially for non-coding work.
On Skills vs project instructions: project instructions are 'who the agent is.' Skills are 'specific things it already knows how to do.' For research and Notion workflows, Skills are probably overkill right now. I'd only reach for Skills when you catch yourself re-explaining the same procedure every time.
awesome read here as per usual.
"Keep your CLAUDE.md thin. Every line you add is a line the model has to read at the start of every single session. Treat it like precious real estate. If you can say it shorter, say it shorter."
100% this. in fact, i think this will be more important during each iteration of the models. two years ago these files needed to be big, and it seems like we can rely on the models to do more of the heavy lifting and reasoning every few months.
if you haven't taken the time to thin out your claude.md and some of your skills, now is a good time to do so as we prepare for the coming drops!
Thanks! and you are right. Instruction file → less is more :D
🔥🔥🔥
Most people optimizing their agent are still in the execution layer — shaving tokens, tuning prompts, but context loading is a different problem. Someone has to decide what gets loaded: you, RAG, or the agent itself, that choice doesn't disappear just because the window is 200k.
true :)
Thanks Pawel. This described some experiments I’ve been building without framing them expressly as agents. Claude Code, Obsidian vault, instruction files with autonomy tiers, scheduled monitors, memory as markdown. I’m about to go deeper with an old Mac Mini that’s currently just sitting idle. Was thinking of trying OpenClaw on it but maybe I’ll build something from scratch for the full learning experience.
Thanks Jonatan. Sounds like you're further along than you think. Autonomy tiers, scheduled monitors, memory as markdown... that's exactly the shape. You've been building an agent without calling it one.
On the Mac Mini question, go for it. The moment I moved off my MacBook was the moment my agent stopped being a hobby project. That old Mac Mini will run a surprising amount of stuff.
OpenClaw vs building from scratch: I'd build from scratch. Not because OpenClaw is bad, but because the rebuild is where most of the learning is. Read other repos for concepts (I was inspired by a bunch of ideas that way), but owning the architecture yourself is how you understand what's yours to tune. Let me know how it goes.
There’s also a third option I’m watching. The signs from Anthropic arguably point toward a built-in agent platform, especially what was visible in the leak. Looks like scaffolding for persistent agents inside the Claude ecosystem. I’m half tempted to wait and see what ships natively before I invest a lot of time in building my own infrastructure.
But that’s a different discussion and the learning experience will probably be worth it regardless of what Anthropic ships next. Starting small like you recommend in your post.
Great article! Thank you so much for your work!!
I tinkered around with Claude projects and I was also very surprised that in the end everything is more or less text. In Claude project you can add instructions (I guess this is the role of the agent), files for context and you have a chat window. I connected Notion to Claude and it created some websites for me according to my design (which was added as a file called brand.md) and transferred a description of the results to Notion. Another experiment I did was an extensive research and then creating Notion databases for me, which worked really well.
In this harness of Claude project, there are no folders on my laptop, it's more the UI of Claude Project, but I did not test yet something that would write into my laptop's folders. So I'm not yet quite sure where in my setup would be the Claude.md.
What I also do find confusing, is the Skills section of Claude. There are many different folders and I do not really get the point of having an agent in the Skills section vs. having project instructions. My questions are related especially to non-coding tasks like researching something for me and creating Notion databases. This all worked without the Skills section in Claude. So I guess this is perhaps relevant later.
Bianca, thank you. You nailed it in your first paragraph: once you realize everything is text, the whole thing stops being mysterious.
On where CLAUDE.md fits in your setup, you're already using the equivalent. In Claude Projects, the 'Custom instructions' field IS your CLAUDE.md. Same idea, different surface. The files you add for context are your project memory. The chat window is your channel. You have the three pieces, just wrapped in a UI instead of on disk. Nothing wrong with that, especially for non-coding work.
On Skills vs project instructions: project instructions are 'who the agent is.' Skills are 'specific things it already knows how to do.' For research and Notion workflows, Skills are probably overkill right now. I'd only reach for Skills when you catch yourself re-explaining the same procedure every time.
Thanks a lot! That confirms exactly what I thought too. No skills needed for non-tech and heavy "language-related" stuff in the first agent setup.