Project Organization Architecture Question #17
Replies: 6 comments 3 replies
|
Hi Joost 👋, Folder StructureI'm curious where you set My understanding of the two project folders:
The context folder is PAI's memory, it looks there first to understand what you are talking about when you mention a project. Then it can understand what the stuff in The Privacy ImplicationsI'm curious how you've set the Git stuff up. Are you using a private GitHub repo to back up your PAI? Are you excluding any specific folders from version control out of privacy concerns? Looking forward to reading about your progress. |
|
This is correct here @joostliebregts. This is how I have things organized. PAI_HOME just accesses my authoritative Claude's settings. I have alieases for: alias cl="bun install -g @anthropic-ai/claude-code; claude --dangerously-skip-permissions" Which launch Claude Code in different directories with different options. But I run CC mostly out of directories with projects, not from home. |
|
Thanks @smolcompute and @danielmiessler for the excellent explanations! @daniel - I should mention I'm not a software developer (industrial design background), though I now focus on using co-creation processes to facilitate more collaborative societal structures, augmented with AI. That's why I'm so fascinated by PAI - it's exactly the kind of human-AI collaboration that could help people do more human things, as you write about. Your aliases explanation was super helpful. If I understand correctly: the "Projects" folder in PAI_HOME is really more of a documentation suggestion than a rigid requirement, right? Since you run Claude Code from wherever your actual project work lives, the key is that your @leo on Git setup - I'm running a more distributed approach, and honestly, Finn explains this better than I do:
Privacy-first approach while still getting the multi-machine benefits! Quick thought on Notion 3.0 - Did you catch their announcement? Interesting to see similar ideas popping up - it seems like they've been inspired by the PAI philosophy (readable sub-agent files, context systems) and wrapped it in a GUI. Could make this accessible to millions of non-technical users, though you lose the portability. Seeing the same patterns emerge across different platforms feels like validation of these concepts. Though I'm a bit worried that the automation (which Notion showed in their keynote) adds a new layer of abstraction, increasing the gap between the actual things we humans connect to each other with. Like Simon Sinek writes about in Leaders Eat Last - when engineers making software did one week of customer service, the amount of valuable fixes for users actually increased. I'm paraphrasing here, and not sure if I'm phrasing this entirely correctly, but the idea is that the closer we are to the actual other person, the more we feel connected. The more steps and complexity in between - you know, the opposite of actually talking to someone and hearing them explain what's bothering them - the easier it becomes to make decisions that aren't emotionally connected to others. So that's what I worry about a little bit with these kinds of automation systems. I'm thinking about what is the more human and connected thing to do when you get AI tools to automate a bunch of the "busy work." Best, |
|
@joostliebregts, thanks for the details on your setup! That’s something I’ll be looking into eventually as an upgrade for sure. Heck, I’d love to run the LLM powering PAI on my own server someday too. One thing at a time! About Notion—yeah, definitely impressive. I’ve stayed away from using it entirely after the hard-learned lesson of depending on a proprietary file format (Evernote). I use Obsidian and I love it, but I know it won’t be getting any of the AI capabilities of Notion anytime soon. The founders actually said that’s not something they’re looking to implement, and I understand why. My hope is that PAI can integrate with my existing Obsidian vault. I completely relate to Daniel’s love of plain text: blazing fast, local, portable, future-proof. I want a system that's entirely mine. But it does require more work. We'll get there! |
|
Yeah, I'm working on Olam integration, and also integrations with other LLMs like OpenAI and such just by using custom commands or custom agents. |
|
Hey Daniel, This seems like a fitting follow-up for this thread, since it's about project organization. I’ve been experimenting with Claude Code on a project I call Social GitHub: a way to document and share which social interventions actually spark bottom-up change (local community solutions → transferable patterns). In my work I facilitate projects and sessions with people who are mostly affected by the rules/systems. The last year I've turned more and more to collaborate with AI to make the sessions more effective, lettings people focus on listening to each other and co-creating, while AI loops back what was said in structured formats. The idea with a Social Github is to scale local wisdom without losing the human agency that makes it meaningful. Here’s the snag: I’ve already tried building this system four times. Each attempt showed promise (it even mapped social networks and ownership patterns from interviews), but the whole thing kept getting too complicated. The turning point came when I discovered PAI. It feels like the missing piece: consistent AI infrastructure instead of reinventing the wheel every time. Your framing of systematic AI use seems to fit exactly what this needs. So my question: how would I set up a separate PAI instance on the same Mac that’s already running Claude Code? I’d like to test PAI for Social GitHub while keeping my current Claude setup intact. This feels very much in line with your Human 3.0 perspective — using AI to amplify human collaboration, not replace it, so I'm curious what you think of this idea too. Warm regards, |
|
Just my guess, but selectively load a different .env file that will simply point to a different ${PAI_DIR}/ ? |
|
Wow, cool stuff. You should check out Substrate! |
|
This has been well answered by the community - thanks @leopoldkristjansson for the excellent explanation! To recap the architecture:
Think of it as: the Projects folder is your work, the context folder is your AI's understanding of that work. Workflow:
Re: running multiple PAI instances - check out Substrate as mentioned. Each instance can have its own context while sharing core PAI infrastructure. Marking this as resolved since the core question has been answered! |
Uh oh!
There was an error while loading. Please reload this page.
Uh oh!
There was an error while loading. Please reload this page.
Background & Thanks
Hi Daniel,
First off, huge thanks for sharing your thoughts on PAI so extensively in your article and YouTube video, and then proceeding to open source the entire system within just two weeks. I was captivated by the initial content and immediately started working with what you shared, figuring things out along the way.
After spending countless hours getting it operational (before your repo release), I'm thrilled to say I have PAI up and running! I've named my personal AI assistant Finn (after much deliberation), and I'd like to introduce him as my AI partner in this journey. Yesterday I implemented the voice server, which makes Finn feel almost like a real person now.. besides that it's genuinely amazing to see how well the universal context system and dynamic context loading work together.
My Question: Project Folder Architecture
As you continue publishing documentation, I'd love to understand your thinking behind the project folder structure. Specifically, I'm seeing references to:
.claudefolder (but within the PAI home directory).claude/context/projects/)In the documentation, these seem to be used in different contexts, but it's not entirely clear to me what your reasoning is for this distinction. I've spent hours getting the system working tremendously well, and as I keep building onto it, I'm particularly curious about:
workingfolder? I picked this up from your article, and I've been using it somewhat interchangeably with project folders, but I'm curious about your specific thinking on this folder's purpose in the architecture.I apologize if I've missed this explanation in your video or article, but I'd love your thoughts on how to properly think about the context system boundaries and organization as I continue developing my setup.
Current Status
The system has been working tremendously well for me, and the more I build onto it, the more I appreciate the thoughtful architecture you've created. Any guidance on the project organization philosophy would be incredibly helpful!
Thanks again for making this available to the community.
Best,
Joost (with Finn)
All reactions