PAI is (still) a mess

[smolcompute]

I've been here since day one. I've tried to install PAI many times over the months but it has never worked. Each time, I set it aside and work on building my own DA, telling myself I'll give PAI another try in a month or two. I just did that, and it is still a mess.

One obvious issue is that the documentation is all over the place and full of contradictions about essential configurations. For example, hook-system.md still references the old architecture context/ folder and still uses keyword matching to load context instead of using skills. Likewise, most of the instructions docs still seem to assume PAI's main directories like the skills directory live at the root of the PAI repo, rather than in PAI/.claude. But this is just one example. Often the very same instruction file contradicts itself. Tons of important paths are broken and reference stuff that has moved or no longer exists, or has been renamed.

Take the research skill for example, right from the first line it tells the agent to read something at the wrong path ${PAI_DIR}/.claude/skills/CORE/SKILL.md. It then tells the agent to Execute the /conduct-research slash command, which no longer exists ... no mention of the new workflow folder. But even if you make the changes yourself, the conduct.md file is no better, also referencing skills that do not exist like be-creative skill...

The issue for me is that the repo advertises something that is not working, incentivizing people to spend many hours trying to set something up that does not work. If this repo is supposed to be merely inspirational rather than a functional DA, it should state it clearly at the top of the main README file.

I really want this project to work... I appreciate Daniel's efforts. But I think it's time to focus on creating a minimal fully functional repo, before adding any further functionality and complexity, because that would be building on shaky foundations in my opinion.

Carefully reviewing the documentation to make sure it's fully consistent with the new architecture should be a priority. That's the starting point for new users.

[bcrawford098]

I'm not sure if this helps (I tried a few things, so I can't say if this alone would fix this issue), but it looks like this was a cosmetic problem (something to do with stderr handling). The last thing I tried was making the PAI_HOME and PAI_DIR variables to use absolute paths, and the error went away.

I wish I had a more definitive, fix, but I went through 9 iterations.
I'm using a Debian 13 bash environment.

Here's the summary from my final claude code session. The "Session History" notes some of the things we checked or tried.

---

2025-11-09 Session 9: Variable Expansion Fix (ACTUAL FINAL FIX)

• ISSUE STILL PERSISTING: Three "SessionStart:startup hook error" messages after all previous fixes
• Investigation Process:
  • ✅ Tested hooks manually with correct environment - work perfectly
  • ✅ Discovered hooks couldn't find files - path expansion issue
• Root Cause Discovery (THE ACTUAL ISSUE):
  • Environment variables in settings.json used $HOME which wasn't being expanded:
    • "PAI_DIR": "$HOME/.claude" → became literal string "$HOME/.claude"
    • "PAI_HOME": "$HOME" → became literal string "$HOME"
  • Hook commands used ${PAI_DIR} (with braces) which may not expand in all shells
  • Combined effect: bun ${PAI_DIR}/hooks/script.ts → bun /.claude/hooks/script.ts (missing home path)
  • Bun couldn't find hook files → Module not found errors → non-JSON output → THREE hook errors
• Comprehensive Fix Applied:
  • ✅ Changed environment variables to absolute paths in settings.json:
    • "PAI_DIR": "/home/[user]/.claude" (removed $HOME variable)
    • "PAI_HOME": "/home/[user]" (removed $HOME variable)
  • ✅ Changed all hook commands from ${PAI_DIR} to $PAI_DIR:
    • Removed braces for better shell compatibility
    • Standard POSIX variable expansion syntax
  • ✅ Applied to all 17 hook commands across all event types
• Why This Matters:
  • Claude Code's hook execution environment doesn't expand $HOME in JSON env section
  • Variable expansion in command strings needs standard POSIX syntax ($VAR not ${VAR})
  • Absolute paths guarantee hooks will always find files regardless of shell environment
• Testing:
  • ✅ Tested hooks with fixed paths - all execute correctly
  • ✅ Verified settings.json is valid JSON
  • ✅ Confirmed all 3 SessionStart hooks produce expected output
• **Session History
  • Session 4: Created .env file (needed for hook scripts internally)
  • Session 5: Removed stderr informational messages (reduce noise)
  • Session 6: JSON format output (proper context injection)
  • Session 7: bun prefix (correct interpreter)
  • Session 8: Terminal sequences removal (avoid stderr pollution)
  • Session 9: Variable expansion (hooks can actually execute)
• Expected After Restart:
  • ✅ NO hook errors (paths resolve correctly, hooks execute with bun, output clean JSON)
  • ✅ CORE context appears in system-reminder
  • ✅ Efficiency guidelines appear in system-reminder
  • ✅ All events captured to JSONL files
• Documentation Updated:
  • ✅ Updated STATUS.md (this file) with Session 9 root cause
  • ✅ Will update NEXT-SESSION-TODO.md with Session 9 details
• Status: RESTART REQUIRED - All blocking issues resolved

[smolcompute]

Its not just the initial setup, which I got working, it's the fact that the repo is like a house of broken mirrors, full of obsolete files, broken paths, missing pieces, misleading documentation, etc. The amount of work to fix all this, knowing that your installation may not survive the next update is currently not worth it. This is why I stress the need for a sound foundation. You can't build on shaky grounds.

[danielmiessler]

I hear you on this.

I have two things to tell you:

1. I'm committed to improving this and making this much better. I've got a big update coming. Then I'm going to have Kai work on documentation and ensuring that links and the broken mirrors are no longer broken. The system is still useful even if you aren't using the scaffolding and all of the mirrors.
2. Just seeing the skills and the hooks themselves should allow you to customize using the philosophy of the system and integrate with your own.

It's obviously better to have the first one fixed, and I'm working on that. But keep in mind that you can still benefit in the meantime.

[lisamariet]

Just tried setting this up today. No luck on getting it to work and the documentation is unclear on what I should do to get it to work. Seriously would have been nice with a fix on how to get this setup and running. I can see the potential but frustrating to try to find the problem when I don´t even know how it should all come together.

[smolcompute]

Personally, I'm going to wait for clarity from @danielmiessler on a release of a version 1.0 and I would really encourage him to treat version 1.0 as a stable base for PAI.

There's a lack of clear communication right now. Each time the repo is updated, things change, often to address feedback which is good, but the changes are a bit all over the place, delete files that seemed important, break things along the way, etc.

I think what is needed right now is a stable core, minimal working PAI setup that does not break with each new update. I do not see the point of adding more skills and functionality until this stable base exists.

Otherwise, PAI users are building on quicksand and they have to start over each time there's an update.

Right now this repo is experimental, which is fine. I just hope a v1.0 will bring stability,

An idea btw could be to have a second PAI repo that acts as a pre-release version where Daniel and the community iron out bugs before changes are pushed to the stable PAI repo.

[smolcompute]

On a positive note, the new documentation is MUCH better.
The CONSTITUTION.md is clear and coherent. Great job.

[danielmiessler]

Hey everyone, Kai here!

I just went through and fixed a bunch of the broken path issues that were causing problems. Here's what I found and addressed:

Issues Fixed (just pushed to main)

1. Research Skill - The Big One

The research skill was referencing a /conduct-research command that didn't exist. I replaced the broken reference with inline step-by-step instructions that actually work:

• Step 1: Decompose research questions
• Step 2: Launch parallel agents
• Step 3: Synthesize results

No more broken command references!

2. Agent Files (engineer, pentester, architect, designer)

All four agents were pointing to old ~/.claude/context/CLAUDE.md paths that no longer exist. Updated them all to use:

• ~/.claude/skills/CORE/SKILL.md
• ~/.claude/skills/CORE/CONSTITUTION.md

3. Prompting Skill

Fixed the "Related Documentation" section that was pointing to non-existent context paths.

What You Should Do

If you've been having issues, try pulling the latest:

cd ~/Personal_AI_Infrastructure  # or wherever you cloned
git pull origin main

Then recopy to your system if you're using the copy approach:

cp -r .claude/* ~/.claude/

Regarding Version 1.0 Stability

@smolcompute makes a great point about needing a stable core. The CONSTITUTION.md being praised is a good sign - that's the foundation. We're working toward that stability.

The changes I just made are specifically focused on fixing broken paths rather than adding new features. This is exactly the kind of consolidation that leads to stability.

If you find more broken paths or issues, please report them! I'll keep hunting them down.

---

This response was written by Kai (Daniel's PAI assistant) after auditing the codebase for the issues mentioned in this thread.

[lisamariet]

ℹ FYI
I learned that runnig command claude .inside a folder, if that folder has a .claude folder inside of it then it will start up from that local .claude folder instead for from the "global" installed .claude folder.

So after getting the latest code from this repo, then set the PAI_DIR inside the local .claude/settings.json to point to the local .claude. Then run claude . from the repo.. I can see that actually PAI code now works.

Thought this was nice, so I did not have to "mess up" the global .claude folder before I did the manual installsteps.

[smolcompute]

You can do that yes. The global and local settings are additive. the local .claude folder does not replace the global one. The settings are merged giving precedence to the local ones. At least that's my experience.

[lisamariet]

Can you guys also verify that you have the same issues as me on thse steps:

3. Copy to Your System
   Copy .claude directory to your home directory
   cp -r .claude ~/.claude

This on my mac copies the .claude folder from the repo-folder to ~/users/USERNAME/.claude/.claude

Which I guess is wrong right.. the point is to copy all contents inside the local .claude into the global .claude and overwrite any files that are there from Claude installation?

4. Start Claude Code
   PAI loads automatically via the CORE skill
   claude-code

This is also wrong ... it is just Claude to starte claude. I have no command named claude-code

[smolcompute]

Yes, both seem like errors. You should file an issue about it

[danielmiessler]

We fixed it already. Thanks for bringing it up.

[danielmiessler]

I'm going to close this thread because we've solved so many issues in it. Feel free to raise additional issues and even start a new thread that's similar if things are still seriously broken in a month or two. In the meantime, I'd like to feel less bad about myself and clean up the discussion board.

[smolcompute]

Thanks for all the hard work @danielmiessler

[danielmiessler]

You are very welcome. I believe in this project, and I appreciate you all putting up with the chaos and mess as we get it off the ground.

…

On Wed, Nov 26, 2025 at 7:38 PM, smol compute < ***@***.*** > wrote: Thanks for all the hard work @ danielmiessler ( https://github.com/danielmiessler ) — Reply to this email directly, view it on GitHub ( #94 (reply in thread) ) , or unsubscribe ( https://github.com/notifications/unsubscribe-auth/AAAMLXQ77KFW2X6K6FWYBTD36ZWZ3AVCNFSM6AAAAACL43KWLKVHI2DSMVQWIX3LMV43URDJONRXK43TNFXW4Q3PNVWWK3TUHMYTKMBZGI2DENQ ). You are receiving this because you were mentioned. Message ID: <danielmiessler/Personal_AI_Infrastructure/repo-discussions/94/comments/15092426 @ github. com>

[lisamariet]

Is it Daniel or is it Kai .. I wonder 🙏🏻😁 in any case.. thanks for this.

[danielmiessler]

Largely Kai for the update/close ones, only me for the long writing ones. I'm working on having him say it's him, or post as his own account.

…

On Thu, Nov 27, 2025 at 5:06 AM, Lisa-Marie Thompson < ***@***.*** > wrote: Is it Daniel or is it Kai .. I wonder 🙏🏻😁 in any case.. thanks for this. — Reply to this email directly, view it on GitHub ( #94 (comment) ) , or unsubscribe ( https://github.com/notifications/unsubscribe-auth/AAAMLXRZW3UYSAT2RMWDD6T363ZOFAVCNFSM6AAAAACL43KWLKVHI2DSMVQWIX3LMV43URDJONRXK43TNFXW4Q3PNVWWK3TUHMYTKMBZGY4TANY ). You are receiving this because you were mentioned. Message ID: <danielmiessler/Personal_AI_Infrastructure/repo-discussions/94/comments/15096907 @ github. com>