🙏Questions that need to be answered.

[evenromo]

I know youre probably a very busy man and have a lot of other things going on in your life. And I dont expect your full attention to this that or the other. Im just very confused about the state of this project and this repo at the moment. I would love to help out, help with the documentation, help with anything really. But at this moment im too confused to be helpful i think haha. Im a big fan of this project and the ideas around it so i want it to succeed. So i would be very glad if you could shed some light on any of this.

Where do you @danielmiessler install your PAI?
While digging around in this repo you stumble upon more and more signs that points to the fact that youre installing your PAI straight in the ~/.claude folder or something similar. Am i correct in these assumptions? Or are there some gotchas that isnt mentioned in the getting started guide? Like @smolcompute asked about when it comes to skills having to live inside a .claude folder. I saw you agreed to his messages that they had to live in those directories. But could you maybe shed some light on why the folders in this repo are structured in a way that is not aligned with Claude Code methodologies?

Have you actually tried to install PAI on a new machine that is not configured to your specs?
Ive tried maybe 15 times to follow the install guides to this repo and i simply cant get it to work properly without putting everything in the ~/.claude folder. I know that im not alone in this experience or experiences closely related to this.

Am i using this repo wrong? Or am i expecting too much? Have I misunderstood something?
From my POV this is a repo that is a boilerplate for building my own PAI. Where you follow the install instructions and the boilerplate should work without moving files or folders around. Am i correct in this assumption? OR should i look at this as a bunch for folders and files to copy and paste to other directories to get started? Im just very confused right now?

[smolcompute]

I too have spent a silly number of hours trying to install PAI without success.

The new documentation is a lot more user friendly in style, but the repo as a whole is very confusing because so much of it is Daniel's own Claude system, which is an opinionated and personalized use of Claude Code. The result is that it's very easy for something to break if you haven't set things up exactly like Daniel has on his own machine.

I believe PAI should instead be a framework that starts from an extremely minimal install that can't break (basically just CC, some environment variables, and some examples of CLAUDE.md files to get you started). Then, the PAI repo should offer a suite of skills that extend this minimal install. Basically plugins for PAI.

Each plugin should come with its own short documentation. As long as the user has set the environment variables in the PAI .env file, the plugins should just work.

I also think doing a fresh install on a test computer account and following exactly the steps in the documentation as it exists right now would be a good experiment for @danielmiessler. It still doesn't work for me.

[danielmiessler]

Hey all, I hear you loud and clear.

Couple things to say.

1. Things are changing so quickly that this is bound to happen, and I'm sorry for the disruption.
2. Although the goal is to have a system that you can just boot up and it just works perfectly, and I do think I'm going to be able to get there soon, you still can get massive benefit from the components themselves.
3. The commands, the skills, the agents, etc. are so useful. Plus being able to see how context is managed with the hooks and everything. Again, I would rather that it just started and worked, and we're getting much closer to that. But the whole system is there still for you to pull from.
4. I've incorporated more documentation in the documentation folder, and things are getting simpler rather than more complex now. This should hopefully be calming down now.

If you haven't pulled the repo recently or looked at the contents, a lot of the issues you're talking about might have already been addressed.

Really appreciate the feedback.

[smolcompute]

Yes, essentially I've used PAI's repo as an opportunity to look into how @danielmiessler sets up his own DA, which is very useful. I've then started building my one on the side, borrowing from Daniel where I can.

I have no doubt that PAI will eventually mature and the issues will be sorted out. It makes sense for the project to be very experimental at this point in time given that the underlying tools are evolving so fast.

I would still reiterate the idea of PAI as a minimalist install, which can then be extended with skills/ plugins shared by the community much like patterns are in Fabric. A PAI marketplace, added via CC plugin, etc. Obsidian, for example, has excelled at this. The community plugins often work together to provide additional capabilities to a single system.

[hjbrandt]

Thanks for bringing up the questions and for your quick response. I have gained a lot from this project and learned a great deal in the process. I got so excited that I overcommitted, even though I know we are not at version 1.0 yet.

I have had everything working a couple of times, but it has taken considerable time and tweaking to make it happen. I will start sharing my tweaks here, so others might benefit (and maybe it could help with documentation).

Keep up the work, and let us know if there are any specific aspects of the project you would like us to test.

[danielmiessler]

[vichong]

Hi all,

I've successfully set up a fork of PAI that can pull Daniel's updates while keeping my customizations. Here's what I learned:

The Solution: Symlinks + Fork Workflow

The core issue: Claude Code needs files in ~/.claude/ but PAI should be a git repo in your projects directory so you can maintain your own fork.

Note: The symlink approach is documented in the Getting Started guide, but the fork workflow (maintaining your own repo while pulling Daniel's updates) isn't clearly explained. That's what this guide addresses.

Working Setup

# 1. Clone PAI to projects directory (NOT ~/.claude/)
cd ~/projects
git clone https://github.com/YOUR-USERNAME/Personal_AI_Infrastructure.git PAI
cd PAI
git remote add upstream https://github.com/danielmiessler/Personal_AI_Infrastructure.git

# 2. Set PAI_DIR environment variable
# Add to ~/.zshrc or ~/.bashrc:
export PAI_DIR="$HOME/projects/PAI"
export PAI_HOME="$HOME"

# Reload shell
source ~/.zshrc  # or ~/.bashrc

# 3. Symlink settings.json from ~/.claude/ to your repo
ln -sf $PAI_DIR/settings.json ~/.claude/settings.json

# 4. Verify
ls -la ~/.claude/settings.json
# Should show symlink pointing to your PAI repo

My Working Structure

~/projects/PAI/                    # Git repo (your fork)
├── .claude/
│   ├── skills/                    # Your machine-specific custom skills (gitignored)
│   └── settings.local.json        # Machine-specific overrides (gitignored)
├── commands/                      # Slash commands (referenced via ${PAI_DIR})
├── skills/                        # Canonical PAI skills
├── settings.json                  # ← Symlinked from ~/.claude/settings.json
└── .env                          # Your API keys (gitignored)

~/.claude/                         # Claude Code's location
└── settings.json -> ~/projects/PAI/settings.json

Note: There are two skills directories:

• ~/projects/PAI/skills/ - PAI's canonical skills (committed to git)
• ~/projects/PAI/.claude/skills/ - Your machine-specific custom skills (gitignored, not synced)

Maintaining Fork-ability

Pull Daniel's updates:

git fetch upstream
git merge upstream/main
git push origin main

Your customizations stay separate:

• .env (gitignored - your API keys)
• .claude/skills/ (gitignored - machine-specific skills)
• .claude/settings.local.json (gitignored - machine-specific overrides)

Quick Verification

# Check symlink works
readlink -f ~/.claude/settings.json
# Should show: /home/user/projects/PAI/settings.json

# Test in Claude Code
echo $PAI_DIR
# Should show: /home/user/projects/PAI

# Test in Claude Code: "What skills do you have?"

---

TL;DR: Clone PAI to a projects directory, set PAI_DIR, symlink settings.json from ~/.claude/ to your repo. Keeps everything in git while satisfying Claude Code's requirements. The fork workflow lets you pull Daniel's updates while maintaining your own customizations.

[smolcompute]

Thanks for this.
However, is Claude automatically loading the skills located here ~/projects/PAI/skills/ ?

I assumed only the skills folders located in a .claude/ folder would automatically loaded into the context window when the session starts. Same for the commands folder.

[evenromo]

This! Last piece of the puzzle!

[vichong]

I wanted to follow up on my earlier response about using symlinks. After working with this setup more, this is how I now understand it works...

1. Global Settings Symlink (at ~/.claude/ level)

~/.claude/settings.json → ~/projects/PAI/settings.json

This single symlink makes your PAI settings globally available to Claude Code.

2. Project-Level Symlinks (within PAI/.claude/)

Within your git repo at ~/projects/PAI/.claude/, you create relative symlinks:

.claude/agents → ../agents
.claude/commands → ../commands
.claude/skills → ../skills

[danielmiessler]

Explain to me again the problem that we are trying to solve. Maybe there is an easier way for us to do this.

…

On Tue, Oct 21, 2025 at 5:25 AM, Victor Hong < ***@***.*** > wrote: I wanted to follow up on my earlier response about using symlinks. After working with this setup more, this is how I now understand it works... 1. Global Settings Symlink (at ~/.claude/ level) ------------------------------------------------ ~/.claude/settings.json → ~/projects/PAI/settings.json This single symlink makes your PAI settings globally available to Claude Code. 2. Project-Level Symlinks (within PAI/.claude/ ) ------------------------------------------------ Within your git repo at ~/projects/PAI/.claude/ , you create relative symlinks: .claude/agents → ../agents .claude/commands → ../commands .claude/skills → ../skills — Reply to this email directly, view it on GitHub ( #52 (reply in thread) ) , or unsubscribe ( https://github.com/notifications/unsubscribe-auth/AAAMLXSA7GD4MXDLVNOIWGT3YYQ3RAVCNFSM6AAAAACJVGKZGCVHI2DSMVQWIX3LMV43URDJONRXK43TNFXW4Q3PNVWWK3TUHMYTINZTHE4TINY ). You are receiving this because you were mentioned. Message ID: <danielmiessler/Personal_AI_Infrastructure/repo-discussions/52/comments/14739947 @ github. com>

[smolcompute]

@danielmiessler the problem is that Claude Code only recognizes commands, agents, and skills when they live inside a .claude/ folder (either the global one or the project's). There is no step in PAI's setup instructions that address this. PAI has these folders live at the root of the repo so they are not recognized by CC. You could add an installation step that tells people to create a symlink for each folder inside their global .claude/ or in PAI's .claude/ .

[danielmiessler]

Or should I just have the repo have.Claude in there one level up, so that if you download the repo, you will have a.Claude directory? Is that easier? I didn't go that route because I wanted to be more agnostic to systems other than Claude Code. But I think you're actually right. Because the interconnectedness of the system is based on Claude Code, then we just have to go with that for now.

…

On Tue, Oct 21, 2025 at 10:06 AM, smol compute < ***@***.*** > wrote: @ danielmiessler ( https://github.com/danielmiessler ) the problem is that Claude Code only recognizes commands, agents, and skills when they live inside a .claude/ folder (either the global one or the project's). There is no step in PAI's setup instructions that address this. PAI has these folders live at the root of the repo so they are not recognized by CC. You could add an installation step that tells people to create a symlink for each folder inside their global .claude/ or in PAI's .claude/ . — Reply to this email directly, view it on GitHub ( #52 (reply in thread) ) , or unsubscribe ( https://github.com/notifications/unsubscribe-auth/AAAMLXU2VH3ZH6EBMBYRCSL3YZR3DAVCNFSM6AAAAACJVGKZGCVHI2DSMVQWIX3LMV43URDJONRXK43TNFXW4Q3PNVWWK3TUHMYTINZUGI3TOMI ). You are receiving this because you were mentioned. Message ID: <danielmiessler/Personal_AI_Infrastructure/repo-discussions/52/comments/14742771 @ github. com>

[smolcompute]

@danielmiessler , yes. I think going CC for now makes sense. Especially given the skills architecture. Also the others don't have hooks, subagents, let alone always-in-context skills. Trying to make PAI work with other agents likely adds complexity and won't really work smoothly anyway.

I vote in favor of a .claude/ folder in PAI's root. This eliminates the need to symlinks altogether actually... And it leaves the possibility of running non-PAI instances of CC elsewhere.

[danielmiessler]

Awesome, we made the change and it's live.

…

On Tue, Oct 21, 2025 at 3:41 PM, smol compute < ***@***.*** > wrote: @ danielmiessler ( https://github.com/danielmiessler ) , yes. I think going CC for now makes sense. Especially given the skills architecture. Also the others don't have hooks, subagents, let along always in context skills. Trying to make PAI work with other agents likely adds complexity and won't really work smoothly anyway. I vote in favor of a .claude/ folder in PAI's root. This eliminates the need to symlinks altogether actually... And it leaves the possibility of running non-PAI instances of CC elsewhere. — Reply to this email directly, view it on GitHub ( #52 (reply in thread) ) , or unsubscribe ( https://github.com/notifications/unsubscribe-auth/AAAMLXUCMSMRIXYUGGTKHQ33Y2ZAHAVCNFSM6AAAAACJVGKZGCVHI2DSMVQWIX3LMV43URDJONRXK43TNFXW4Q3PNVWWK3TUHMYTINZUGUYDANY ). You are receiving this because you were mentioned. Message ID: <danielmiessler/Personal_AI_Infrastructure/repo-discussions/52/comments/14745007 @ github. com>

[evenromo]

HUUGE!! Awesome work everybody.