Replies: 19 comments 25 replies
-
|
Not once have i actually been able to follow the guide on how to install PAI and have it successfully come out on the other end. This repo is slowly becoming a perfect example for the people that is saying "dont use AI in production". |
Beta Was this translation helpful? Give feedback.
-
|
Dude. I get it. I accept all the criticism. I'm working on it. But try to be nice about it please. It'll get better. I'm working on it. By myself. If you don't think you can do that, just go find another project. There's plenty out there. :) |
Beta Was this translation helpful? Give feedback.
-
|
@evenromo Wow, easy tiger. Tell your AI that wrote this discussion post to have a little grace, please. |
Beta Was this translation helpful? Give feedback.
-
|
I hear the frustration @evenromo, and the issues raised here are real. But I’d like to shift the tone toward a more constructive discussion: what can the community do to help, and what would make a meaningful difference to PAI’s development? I’ve experienced similar installation issues but managed to work my way around them to successfully install and run PAI on my system. I’m on macOS though and I haven’t tried Windows or Linux. I’ve spent a lot of time over the last couple of months expanding vanilla PAI with my own skills and also attempted to contribute some of them back to the community. One of my motivations has been thinking long-term: making my system reinstallable and maintainable over time. It’s quite easy to end up in a place where your own PAI becomes complex to reinstall from scratch. What I discovered during my efforts to create skills that are both compatible with vanilla PAI and publicly shareable is that it’s genuinely tricky to unbolt your work from your personal environment into something portable. At some point I decided to be a “0.1x developer” and invest in TDD tooling with test suites for unit, CLI, integration, and acceptance testing. I also developed a release process with approval gates to create a clear pathway from development to public PR. I think the project would benefit from some tooling around the infrastructure and release process. For example: how can we simplify installation testing for contributors? Do we need separate Docker containers or VMs to test fresh installations? From my own experience, it’s tricky to develop on the same machine as your working PAI. Another learning from the last couple of months has been discovering OpenSpec and SpecFirst principles. I’ve used that as a foundation for all my PAI skill work, it simplifies testing and provides a clear record of what’s actually included in each release. |
Beta Was this translation helpful? Give feedback.
-
|
I've installed PAI from scratch on both my laptop and main development desktop. It took some doing. I hit pathing issues and ran into several of the problems @evenromo raised. But by using Claude Code to work through the installation, I resolved them. Is this an installation I'd hand to someone without a technical background? Absolutely not. But I view it from a different perspective. We're living in the era of non-deterministic software. AI enables bespoke systems that do exactly what you need. PAI isn't a product; it's a foundation you build on. Some of it may not work on every system, and that's what makes it unique. Something doesn't work on Linux? Create a workaround, build an alternative, or skip it. Work with your coding agent to solve that challenge your way. This isn't shrink-wrapped software where you double-click an installer and everything works. One of the best things about PAI has been the exploration: what does it do? What can it help me with? How does it work? To get the most out of these systems, we need to build on them ourselves. The first thing I did after setup was build custom agent personas for my business. The system has exceeded my expectations. I haven't fully transitioned to it as my main environment yet, but I've already unlocked new workflows, especially around agents for iterative design and Gemini nano banana pro integration. The complexity isn't a bug; it's the reality of where we are with this technology. Expect to roll up your sleeves. It's messy. We're figuring it out as we go. And often, there's no single right answer, just your answer. |
Beta Was this translation helpful? Give feedback.
-
|
Sadly I've learned for myself that you are better off when you build your own framework from the very beginning. So you can build it as you like and you don't need to hope a maintainer accepts your PR after you spend hours of developing and testing (more on the later). Furthermore you don't need to be scarry that your provided features are screwed up in the next MINOR or PATCH (NOT MAJOR) Version (e.g. the whole Voice Server Linux & Google stuff). PRs and issues are closed unresolved directly with every new version. Thats not how you interact with a community nor how you should develop software or any other product. Not to say the history or now called memory system (README still talks about history). It never worked for me. Others and I opened multiple issues which are never got resolved. Maybe I come back later when this project matured. Right now its wasting hours to set it up and get it to work more or less. This is mainly because path references are wrong or features not fully deployed (missing files etc.). One suggestion for the future: Let the community help to make a stable version and implement changes via PR not trunk based. Therefore the community can test changes and the changes are merged after community gave feedback. This will prevent from pushing potential breaking changes with every commit right into main. If its to time consuming for one person to maintain its a good idea to invite some additional maintainers to help getting this the right way. EDIT: Or at minimum use a dev branch and merge after fully tested. I really hope this project got on track somewhen because the idea itself is brilliant. |
Beta Was this translation helpful? Give feedback.
-
|
Could not have said it better myself! |
Beta Was this translation helpful? Give feedback.
-
|
Yes, I really like your point about letting you guys help me. Please tell me how I can be better at doing that. I guess the obvious answer is doing what I talked about in the long comment below and basically paying closer attention to PRs and issues and making sure they don't get overwritten. I'm going to try to make that a priority. I do worry about how fast things are changing though, and how fast this idea is evolving for me. Another idea I have here is basically having two branches:
And then the other version is where I'm over here being crazy and doing lots of updates on the other branch. If anyone wants to help me with this, I would appreciate it. That way we can have the best of both mindsets. |
Beta Was this translation helpful? Give feedback.
-
|
I think that there's really three different areas to tackle here. And I do think it's really important to nail some of this down in order for a good community to build. First, the vision can be a bit foggy sometimes. I think a lot of that is due to, as you pointed out, just how fast this stuff evolves, even in the course of development. I find that to be a problem for myself in my local PAI install as well, and I don't have to publish anything. Making stuff public is a whole different level of scrutiny. So having a grand 'vision' is important. Especially with things like 'AI does the installs' and how it conflicts with being deterministic at first glance, but a second look makes it obvious WHY you think that's the way. Second, having two or three branches would be a great idea. I'm not really well versed in the ins and outs of managing an open source project but it seems pretty clear that having one branch that gets dumped on with updates with little notice is kinda rough. A stable branch, a 'stabilizing' branch, and a 'skunkworks' branch might make sense. Skunkworks being where all the new stuff goes to get vetted by the brave. How all this is implemented, well, I hope someone here has some experience there, because I have no idea! Third, the upgrade paths are brittle. I love the new SYSTEM v USER difference in CORE (at least in theory, I haven't updated yet) because a lot of my work is in some of the core skills (Art in particular) and for the last upgrade I actually had to create a new skill to upgrade packs w/o killing my changes. I think that 'System v User' concept should be a core part of the Pack requirement. I'll probably have a lot more to say about this once I do the upgrade. All that said, it's important that you know how grateful I am for the work you've done. And I am confident I'm not the only one. The things I've been able to do since adopting PAI as a foundation for my work is mind boggling, and I can't wait to help improve it. I own a couple companies and I intend to adapt this for each of them as well, and ideally teach other folks in my market how to do that. This stuff is too important to stuff behind a paywall. |
Beta Was this translation helpful? Give feedback.
-
|
@lifegenieai @mellanon @danielmiessler Before i continue... I want to make it clear that i love the IDEA of this system and i love tinkering and figuring out how things work and fixing issues that shows up along the way. Ive actively used and observed the issues, PRs and discussions in this repo almost every day since august and installed multiple versions of PAI multiple times. The best version, that actually worked and did what it promised, was the last version before Anthropic launched their Skills. After that, this repo has been in steady decline. My claims are based on what ive observed since august, and not some "i just got here and i cant seem to be able to install this" tantrum. Ive probably been here longer than most.
I'll silently continue to observe this repo in the coming months but as of right now this repo is a shitbox without a rudder or a captain. Rant over |
Beta Was this translation helpful? Give feedback.
-
|
That said. I personally figured out the 2.0.0 pack version (and some additional commits) worked best for me. The CORE skill was simple and clean. Now we've got a 2.1.0 with bloated CORE skill. E.g. DA configuration, in 2.0.0 a simple configuration inside It does not make sense to me why we now configure markdown files instead of setting technical values by ENV (e.g. default voice id, names etc.) and the read it inside the hook. Thats not error proven because But this shows the root cause of the problem. Every new idea gets pushed to main directly without verifying the consequences. Like some said it works exactly in one MacOS system and some of these features are merged to this project regardless what breaks. Sadly this feels like vibe coding the whole system. Quick wins:
|
Beta Was this translation helpful? Give feedback.
-
|
Thanks for saying that, @evenromo . I hear you loud and clear. |
Beta Was this translation helpful? Give feedback.
-
|
Hey my friend.
So, first off. I am about to move all these definitions to settings.json because that's way smarter and we already have some there anyway, set by the wizard. Great idea thanks.
Second: don't be an ass. Yes, the project isn't professional yet. Yes it's annoying. I'm sorry. I apologize.
It's just me and this whole AI thing is happening at the speed of light man.
I'll do better. But don't bring negative energy to this project.
I'm cool with you punching me in the face. But only as a coach. Not as an internet troll.
I'll listen without the vitriol.
…On Sat, Jan 10, 2026 at 10:16, Timo S. < ***@***.*** > wrote:
That said. I personally figured out the 2.0.0 pack version (and some
additional commits) worked best for me. The CORE skill was simple and
clean. Now we've got a 2.1.0 with bloated CORE skill. E.g. DA
configuration, in 2.0.0 a simple configuration inside SKILL.md and ENV. Now
its a additional DAIDENTITY.md which got not touched by install process but
the hooks now read out of it instead using the ENV variables which where
set in the installation process (name etc). Furthermore DAIDENTITY.md and SKILL.md
have duplicate configuration like personality calibration.
It does not make sense to me why we now configure markdown files instead
of setting technical values by ENV (e.g. default voice id, names etc.) and
the read it inside the hook. Thats not error proven because DAIDENTITY.md is
a the so called "user-scope" which everyone can and should modify. This
might break hooks in the long run.
But this shows the root cause of the problem. Every new idea gets pushed
to main directly without verifying the consequences. Like some said it
works exactly in one MacOS system and some of these features are merged to
this project regardless what breaks. Sadly this feels like vibe coding the
whole system.
Quick wins:
* Commit new changes with PR or at least a dev branch.
* Talk to the community
* Release only stable and tested versions
—
Reply to this email directly, view it on GitHub (
#361 (reply in thread)
) , or unsubscribe (
https://github.com/notifications/unsubscribe-auth/AAAMLXUPCG3NQKQVU6JEBRT4GE6WBAVCNFSM6AAAAACRG2G2BWVHI2DSMVQWIX3LMV43URDJONRXK43TNFXW4Q3PNVWWK3TUHMYTKNBWGQZTENY
).
You are receiving this because you were mentioned. Message ID: <danielmiessler/Personal_AI_Infrastructure/repo-discussions/361/comments/15464327
@ github. com>
|
Beta Was this translation helpful? Give feedback.
-
|
First off, I'm not a native speaker so I appologize if something was to offend. It was not in my mind to punch you in the face. The vision of the project is great and I'm glad if I can help you out if you want to. I think I showed this with the various contributed PR last week. But then I was very frustrated after everything changes after a "major release" (2.0.0) and all PRs were closed. If there is something to discuss I offer you my sparring. |
Beta Was this translation helpful? Give feedback.
-
|
No problem, all is good. Thank you. |
Beta Was this translation helpful? Give feedback.
-
|
Hey everyone. This last CORE change where I separated USER and SYSTEM is the biggest change I've made in a long time, and I think this should be the last major shift like that for a while. The next major one will be around making The Algorithm the primary method of executing things, but that won't change the CORE file system much. Someone mentioned not running this in production. What are you talking about. This is a personal AI stack. This is like your personal agent stack, for getting your context into agentic activities you're doing. I don't even know how you'd run this in production. But if you mean as your daily driver then I hear you. Point taken. As for the install being AI based, that's not because the project is in flux. It's how I think all software will be installed soon. Nobody knows your system better than an AI that can fully inventory it before installing something, and look at your goals and configs, etc., and decide what--if anything--should be changed. I think that's just a smarter way to do installs in a world where things are quite custom and different for everyone. I also want to address the closing of Issues and PRs. We attempted to do that for only things that would be handled properly by the new changes going forward. Meaning, they would be moot afterwards and we could resubmit them if they ever came up again. It was not my intention to just close out every single thing that people had put work into when it wasn't related to the stuff that was being updated. So if that happened, then I'm sorry about that. I will be much more careful in the future to store those or keep them alive and then do proper integrations. The other problem I have in that area is doing an integration, and then having a future change undo that. So I need a solution for that as well. Another thing that I implemented or started implementing in 2.1 is the idea of the user section of core extending the system section of core. Which means any personalization will go into the USER area vs. the SYSTEM or Skills themselves, to minimize overwrites. I'm trying to get to an obvious state, which is being able to update pieces of core, install new Packs, and upgrade functionality without disrupting the customization that people have done to their own PAI systems. I know this has been annoying, but I'm telling you it's going to be worth it. I'm trying to get some additional help on the project from someone who has lots of experience with production maintenance of codebases. Things have been in massive flux just because I was trying to figure out what the project was and how to set it up. But I think a lot more of that is locked down now, so the disruption should definitely slow down. I really appreciate you guys maintaining a positive energy around this. Feel free to keep bashing me when I make mistakes. But please try to do it in a constructive way so we can actually continue making progress without destroying all the hard work you guys have spent on your configurations. Thanks for being part of this. |
Beta Was this translation helpful? Give feedback.
-
|
My gut feel teels me that it would be a good idea to extend the project with evals... This would mean go slow for a bit to stabilise but will mean shipping more confidently in the long run. This can be a good place to start. https://www.anthropic.com/engineering/demystifying-evals-for-ai-agents This could be a parallel work stream where the community can help? |
Beta Was this translation helpful? Give feedback.
-
|
Super ironic since I'm building a whole bunch of eval and test-driven development into PAI itself. Yet the PAI project itself doesn't have it yet. : ( |
Beta Was this translation helpful? Give feedback.
-
|
I have learned a bunch from this repo and the related videos. Thanks.
I too have found the install with AI does a number of things that discourage using this repo. I would really appreciate a single programmed and tested install script that does the install reliably. The current install.ts one in my multiple testing does not do this. AI Install:
|
Beta Was this translation helpful? Give feedback.
-
|
⬆️ THIS! |
Beta Was this translation helpful? Give feedback.
-
|
I agree, I’d prefer a deterministic installation process. The AI can orchestrate and run CLI commands to execute installation steps, but I don’t think having code in markdown files is a good approach, at least not across the board. It adds complexity and introduces another step when moving from development to production, which requires additional validation and QA. |
Beta Was this translation helpful? Give feedback.
-
|
I'm about to update my eval's skill with a whole system built around this. I've been wanting to upgrade this whole system for a long time and now we have some really good guidance here from Anthropic.
Ultimately, I'm looking to have the algorithm be able to wield to this system, as well as all the other systems within PAI.
https://www.anthropic.com/engineering/demystifying-evals-for-ai-agents
…On Sat, Jan 10, 2026 at 14:11, Chronos Enterprises < ***@***.*** > wrote:
I think that there's really three different areas to tackle here. And I do
think it's really important to nail some of this down in order for a good
community to build.
First, the vision can be a bit foggy sometimes. I think a lot of that is
due to, as you pointed out, just how fast this stuff evolves, even in the
course of development. I find that to be a problem for myself in my local
PAI install as well, and I don't have to publish anything. Making stuff
public is a whole different level of scrutiny. So having a grand 'vision'
is important. Especially with things like 'AI does the installs' and how
it conflicts with being deterministic at first glance, but a second look
makes it obvious WHY you think that's the way.
Second, having two or three branches would be a great idea. I'm not really
well versed in the ins and outs of managing an open source project but it
seems pretty clear that having one branch that gets dumped on with updates
with little notice is kinda rough. A stable branch, a 'stabilizing'
branch, and a 'skunkworks' branch might make sense. Skunkworks being where
all the new stuff goes to get vetted by the brave. How all this is
implemented, well, I hope someone here has some experience there, because
I have no idea!
Third, the upgrade paths are brittle. I love the new SYSTEM v USER
difference in CORE (at least in theory, I haven't updated yet) because a
lot of my work is in some of the core skills (Art in particular) and for
the last upgrade I actually had to create a new skill to upgrade packs w/o
killing my changes. I think that 'System v User' concept should be a core
part of the Pack requirement. I'll probably have a lot more to say about
this once I do the upgrade.
All that said, it's important that you know how grateful I am for the work
you've done. And I am confident I'm not the only one. The things I've been
able to do since adopting PAI as a foundation for my work is mind
boggling, and I can't wait to help improve it. I own a couple companies
and I intend to adapt this for each of them as well, and ideally teach
other folks in my market how to do that. This stuff is too important to
stuff behind a paywall.
—
Reply to this email directly, view it on GitHub (
#361 (reply in thread)
) , or unsubscribe (
https://github.com/notifications/unsubscribe-auth/AAAMLXT4IOM3IANBMLBZDKD4GF2HVAVCNFSM6AAAAACRG2G2BWVHI2DSMVQWIX3LMV43URDJONRXK43TNFXW4Q3PNVWWK3TUHMYTKNBWGUZDMMY
).
You are receiving this because you were mentioned. Message ID: <danielmiessler/Personal_AI_Infrastructure/repo-discussions/361/comments/15465263
@ github. com>
|
Beta Was this translation helpful? Give feedback.
-
|
@danielmiessler
I get this, i see your heroic effort to keep up with Anthropic, and youve done an amazing job for the most part. One thing i would like to point out though is; I dont agree with the first part of the sentence ("Its just me"). That simply isnt true, many of the people that has posted, discussed and complained in these channels, including me, truely wants to help. But it does not seem, from my perspective, that youre letting us help. I bet it will be a shitshow initially if you accept others to help but my point still stand, youre not alone in wanting to see this project succeed. |
Beta Was this translation helpful? Give feedback.
-
|
Dude, this seriously makes me so happy to read. 100%, no hard feelings. I also know how it feels to be excited about something and to have it fucking go to shit right in front of you. So I get the frustration. I can't guarantee that this won't happen more, but I'm seriously trying to minimize it, and I think we're going to get to a good place soon. Thank you for being awesome. |
Beta Was this translation helpful? Give feedback.
-
|
Just coming in to say I'm with the sentiment of frustration. I've faced it as well since I haven't been spending dedicated time to get it all working the way I want - but at the same time it's let me see Birds Eye as I start fresh often what is/n't working. This educates me on the framework and how to effectively understand what it is vs just clicking a button and saying cool I have this thing working. With a career in tech, this has always been my valuable side is learning the things vs just being happy I can click and it works. I like breaking shit to find the right ways and maybe even better ways! That said - I did have 2 things to share/ask:
PS - Keep up the good fight. I love this repo and what you're designing. I'm hoping I can get to the point where this is my daily driver for a lot of things. Seen all your usage in podcasts, shorts, clips, and even your blog. Love the convos you were having with NetworkChuck! And hopefully I get to attend a con and meet! Hacker summer camp 2026? 😉 |
Beta Was this translation helpful? Give feedback.
-
|
There is only so much I can decouple from Claude Code without making the system not usable. I think I'm pretty far along that path already. This next 2.2 update will have a lot more cleanup, but it will still be a Claude Code system underneath. Yeah, I think this is a cool idea about work vs. personal. The 2.2 update is going to separate user vs. system first, but I like the idea of work vs. personal as well. For private convos, I'm looking at incorporating local models via routing soon. Using new tools is exactly why AI installs, I think, are really good because you could simply have your AI look at the skills or the hooks or the voice or whatever and say, "Hey, implement this in Gemini." |
Beta Was this translation helpful? Give feedback.
-
|
Yes it's an interesting thought too about toeing the line between making it as deterministic as possible, while also allowing it to "freestyle" it a bit. Because really we need a perfect blend of the two for this scenario. I don't want it to follow everything to a T blindly, this results in many of what I think breaks in the attempts thus far. But we don't want it to just go off the rails doing some wacky things that make no sense. We need to let it have a brain - stealing your terminology in the design principle: let it understand how to build the scaffolding, rather than handing it a specific set of pieces and saying THIS is the only way to build the scaffolding. Let the AI be the General Contractor, using the blueprint as the overall picture of how to build it but letting them recognize when things need altering/changing to produce the end result as the blueprint intended. |
Beta Was this translation helpful? Give feedback.
-
|
@danielmiessler I'd say that decoupling from Claude Code is actually not a great direction. It adds a ton of complexity for the edge case of a minority of users who will always struggle using the system in ways it isn't really designed for. Instead, consider a pattern of embracing Claude Code as an orchestrator. Very similar to how the Art skill works, a PAI skill can decide where to send things. Use gemini for X, notebookLM for Y, send a dev plan to gemini, codex, and grok and build a comprehensive plan based on review, etc. |
Beta Was this translation helpful? Give feedback.
-
|
@ecielam I don't agree at at with the take that decoupling from Claude code is for a minority of users. On the contrary, running PAI locally is a must if this should be a true personal assistant. @danielmiessler I'm very much looking forward to your local version, let me know if you want me to test things eg. on Asus GX10 |
Beta Was this translation helpful? Give feedback.
-
|
Just to give you an example on the chaos that is happening when trying to install this. Look at the time.... 15 minutes. And that is ONLY for the four tasks. Is it the fact that im running Linux that is the issue? Is this install process working seamlessly on MAC? I wish i could just download a zip file, follow some copy + paste instructions, run a couple of terminal commands and have it up and runnning. I think its a mistake that PAI is currently trying to be AI-provider agnostic, i think we should use Anthropics plugin feature. Make this work for ONE provider and then expand or have someone else port it to whatever provider they would want to use. With the current repo structure its so confusing and impossible to know what is happening without being a detective and reading 400 different markdown files and guess which one is the one that is the newest one and the one i should follow. Right now theres no way to know if youre installing it correctly or not. And i have no idea if any of the bits are working correctly because we havent been shown how it is supposed to be working and there are no documentation on how any of the bits and pieces are supposed to be used and work. And since its the AI that has done all of the installing i have no ownership and knowledge about anything, its just automagically supposed to work. But the install instructions hasnt worked once. Ive asked this before, and i will ask again. Are you actually testing the install on a fresh machine or anything like that? Please i BEG you. This does not have to be so complicated. Its markdown files and JS files. A intricate install procedure is just completely unnecessary. Just give us a guide on how to copy and paste the files. I just spent 40 minutes trying to have claude install PAI and it failed AGAIN. If this is the future way of installing software, i dont want ANY part of it. Deleting this for the 1000th time. |
Beta Was this translation helpful? Give feedback.
-
|
I hear what you're saying, but can't you just ask your AI what the current state of your system is and what the state of the PAI pack is and how to unify them? Large overriding opaque installs are also very hard to understand, and it's just a matter of belief. As for not knowing what the AI did, I mean you still have the install script, right? You could still read that or have the AI explain the install script. I feel like it's all still there, the same way, actually better than if it was a brute force install. As for the testing, yes, 100% agree with you. What I'm looking to do is have two branches next week or the week after where we have a stable branch and maybe even a fresh full install that somebody could use as a starting point for every release or at least every few releases. Really sorry. This is causing so much frustration for you. I'm working on it. |
Beta Was this translation helpful? Give feedback.
-
|
The current state right before attempting any of my last attempts is a 100% clean install. All claude/PAI related files and directories are deleted before doing exactly what the documentation tells me to do. I even reinstalled claude code it self before attempting the install |
Beta Was this translation helpful? Give feedback.
-
|
Plugins are static and hard to change. So that's not really a good route for a 'working' system, since by very nature of the system we're going to customize it to our needs, taking it out of plugin territory. |
Beta Was this translation helpful? Give feedback.
-
|
Doing 2.2 tonight that should be much better. Tons of cleaning of Kai/Daniel personalization too.
…On Sun, Jan 11, 2026 at 10:43, evenromo < ***@***.*** > wrote:
⬆️ THIS!
—
Reply to this email directly, view it on GitHub (
#361 (reply in thread)
) , or unsubscribe (
https://github.com/notifications/unsubscribe-auth/AAAMLXSASYNIA5GYMQUJTNL4GKKWNAVCNFSM6AAAAACRG2G2BWVHI2DSMVQWIX3LMV43URDJONRXK43TNFXW4Q3PNVWWK3TUHMYTKNBWHE4DGMQ
).
You are receiving this because you were mentioned. Message ID: <danielmiessler/Personal_AI_Infrastructure/repo-discussions/361/comments/15469832
@ github. com>
|
Beta Was this translation helpful? Give feedback.
-
|
But that means overwriting what you've changed since, right?
…On Sun, Jan 11, 2026 at 13:32, mellanon < ***@***.*** > wrote:
I agree, I’d prefer a deterministic installation process. The AI can
orchestrate and run CLI commands to execute installation steps, but I
don’t think having code in markdown files is a good approach. It adds
complexity and introduces another step when moving from development to
production, which requires additional validation and QA.
I think we’ll see packs and bundles with complex software components that
are better served by traditional installation methods, so installation
will need to support all of the above. That said, I can see
non-deterministic approaches adding value when doing things like merging
code from packs into core components like CORE/skill.md.
—
Reply to this email directly, view it on GitHub (
#361 (reply in thread)
) , or unsubscribe (
https://github.com/notifications/unsubscribe-auth/AAAMLXVLEGTLTLFHLHAKRIL4GK6PDAVCNFSM6AAAAACRG2G2BWVHI2DSMVQWIX3LMV43URDJONRXK43TNFXW4Q3PNVWWK3TUHMYTKNBXGA2TMMI
).
You are receiving this because you were mentioned. Message ID: <danielmiessler/Personal_AI_Infrastructure/repo-discussions/361/comments/15470561
@ github. com>
|
Beta Was this translation helpful? Give feedback.
-
|
Ok I'm starting work on a full release .claude directory that's sanitized and templated like earlier releases. This way users will be able to use one or both of the installs: full releases, or PAI Packs to add on top of them. The new separation of user and system should help this where hooks, skills, and such can be upgraded in place without overwriting customizations that people have made. However, you should always have a backup before starting any of these installs, and plan on bringing over customizations after the install in case that's necessary. But I am listening, so I guess I'm going to offer both releases and Packs as install options. Thank you all for the feedback. |
Beta Was this translation helpful? Give feedback.
-
|
First and foremost, I just wanted to say that I think the system is really cool and shows a lot of promise and I'm really excited for where this could go. All of this progress is really exciting. I installed during 2.0, and haven't upgraded to 2.1 yet because I've made some pretty significant modifications to my PAI system, including external observability and metrics tracking using Datadog, as well as several of the skills. I'm excited about the new user vs system separation and excited to port over my customizations to that instead of the core system itself, but I'm going to wait on a couple of the things discussed in this thread before I follow through on that. I spent a little bit too much time over the past week working on getting this functional in my system instead of things I should be doing elsewhere. I'm super excited to see where this goes and I'm going to be following very closely. Might even try and jump in on some of the discussions given an opportunity. Keep up the great work! |
Beta Was this translation helpful? Give feedback.
-
|
A note of appreciation amidst the frustration I've been following this thread and want to offer a different perspective. |
Beta Was this translation helpful? Give feedback.
-
|
What about making PAI a CLI tool like With PAI as CLI you can integrate hooks: You can apply unit tests on it to make sure nothing breaks and all works as expected. For skills and stuff there could be Handlebar templates where the path issues get resolved based on the users system.
Or or
Tools used in skills should also be included and unit tested within the CLI package. Furthermore there could be something like a "plugin" system which then allows the community to contribute plugins which may only work on one platform and therefore not integrated in the official PAI system. Community plugins are hosted in different repos. This is a very raw idea written in minutes but this would solve multiple problems and accomplish multiple PAI principles. |
Beta Was this translation helpful? Give feedback.
-
|
Do you mean broadly replace everything with a command based system, or augment what's in place with a command based system that does certain deterministic things that are 'generally useful'? |
Beta Was this translation helpful? Give feedback.
-
|
Really interesting! 🧐
…On Sun, Jan 11, 2026 at 23:11, Timo S. < ***@***.*** > wrote:
What about making PAI a CLI tool like fabric ?
With PAI as CLI you can integrate hooks: pai hook xxx
and things like voice/notification server and statusbar.
You can apply unit tests on it to make sure nothing breaks and all works
as expected.
For skills and stuff there could be Handlebar templates where the path
issues get resolved based on the users system.
You then could do something like:
pai skill install core -> installs SKILL.md etc. based on the Handlebar
templates but incorporates platform specific or user specific (e.g. DA
name).
Or
pai skill upgrade --check -> based on frontmatter versioning we can easily
identify changed skills.
or
pai skill upgrade core -> which then writes to a temporary directory and
gives you a detailed prompt (through Handlebar) to give to your favorite
AI to integrate the changes to your system. Maybe incorporate the UserAskTool
for easy migration.
Tools used in skills should also be included and unit tested within the
CLI package.
Furthermore there could be something like a "plugin" system which then
allows the community to contribute plugins which may only work on one
platform and therefore not integrated in the official PAI system.
This is a very raw idea written in minutes but this would solve multiple
problems and accomplish multiple PAI principles.
—
Reply to this email directly, view it on GitHub (
#361 (comment)
) , or unsubscribe (
https://github.com/notifications/unsubscribe-auth/AAAMLXUIK4KVH5SRQEXTTLT4GNCJ5AVCNFSM6AAAAACRG2G2BWVHI2DSMVQWIX3LMV43URDJONRXK43TNFXW4Q3PNVWWK3TUHMYTKNBXGI2TINA
).
You are receiving this because you were mentioned. Message ID: <danielmiessler/Personal_AI_Infrastructure/repo-discussions/361/comments/15472544
@ github. com>
|
Beta Was this translation helpful? Give feedback.
-
|
I understand the frustrations, but this is a new project that is fast gaining momentum so lots of frustrating and breaking changes will likely occur until it matures to a point it will stabilise. As Ive been developing a few packs over the last week and currently in as PR I see these changes and quickly realise this project will become hard to maintain as the momentum grows, especially around third party packs. so I asked Claude to put his thinking cap on :D I have just posted a proposal in the ideas section for a plugin architecture to help de-centralize things, but mainly to start the discussion around the pack architecture, which I hope, will ultimately help ease the burden on the project maintainers as a whole and let them focus on the core functionality and stability. |
Beta Was this translation helpful? Give feedback.
-
|
This is fair criticism and I want to address it directly. The 'giant Jenga tower' problem was real — v1.x was brittle, the install was fragile, and breaking changes came too fast without migration paths. v3.0 was built specifically to fix this: a proper installer ( |
Beta Was this translation helpful? Give feedback.
Uh oh!
There was an error while loading. Please reload this page.
-
PAI installation struggles reveal systemic complexity problems
The Personal AI Infrastructure (PAI) repository has experienced persistent, documented installation difficulties since August 2025. The maintainer Daniel Miessler himself acknowledged the core problem in December 2025, describing the v1.x architecture as "a giant Jenga tower" that "didn't work" — a candid admission that validates user frustrations. Despite an architectural overhaul to v2.0, critical installation issues remain unresolved, particularly for Linux and Windows users.
Maintainer acknowledged fundamental design failure
The most compelling evidence of installation problems comes from the project's creator. In GitHub Discussion #222 (December 20, 2025), Miessler stated:
He elaborated further: "The entire Claude Code system and therefore the entire Kai System is a giant Jenga tower of interconnected files and components. It's a miracle that Claude Code itself ever loads for anyone, anywhere." This frank acknowledgment reveals that the installation struggles reported by users weren't edge cases—they were symptoms of architectural failure that the maintainer struggled to address for months.
The project has undergone four breaking structural changes since its inception: v0.4.0 restructured from nested
PAI_DIRECTORY/to flat structure; v0.6.0 moved all configuration to.claude/directory; v2.0.0 transitioned to modular Packs; and v2.1.0 introduced directory-based pack structure. Each change required users to reconfigure their installations, with documentation frequently lagging behind actual requirements.Linux users face critical case-sensitivity breakage
Multiple open issues document that PAI is fundamentally broken on Linux systems due to case sensitivity mismatches throughout the codebase:
raw-outputsin lowercase, but the actual directory isRaw-Outputswith TitleCase. This works silently on case-insensitive macOS but completely fails on Linux.statusline-command.shuses macOS-onlystatsyntax that breaks on Linux systems.workflows/claude-research.md) while actual files use TitleCase (workflows/ClaudeResearch.md), causing workflow routing to be "completely broken."Community tester @hjbrandt reported: "We just spent a day QA'ing the current Skills system and found TitleCase inconsistencies across nearly every skill." These aren't minor inconveniences—they represent fundamental incompatibilities that prevent Linux users from using PAI at all, yet many remain unresolved weeks after being reported.
Fresh installations failing completely
Perhaps most damning is Issue #224 (December 20, 2025), titled simply "Fresh setup and nothing works." The title itself captures widespread user sentiment. This issue indicates that even following official documentation from scratch results in complete installation failure.
GitHub Discussion #114 documents another typical experience where user GaryGealy reported getting "a ton of hook errors" after installation, including multiple file-not-found errors for essential hooks. Their frustration was palpable: "I really don't have the patience to deal with this ATM." The community—not maintainers—provided the workaround: delete the
~/.claudefolder entirely and start over.Windows users face even greater challenges. User @Gerkinfeltser reported in Discussion #222: "Hooks really seem to be the most tricky & sticky (mostly due to platform differences I am battling on Windows)." The official documentation lists Windows support as "compatible with minor tweaks" and "future," effectively leaving Windows users unsupported despite claims of cross-platform compatibility.
Installation complexity requires AI assistance to manage
The official README now opens with a striking admission: "Manually reading READMEs is pre-2026 behavior. In 2026+, software gets installed by your AI assistant." While framed as forward-thinking, this philosophy shift implicitly acknowledges that the installation is too complex for humans to reliably execute.
The current installation requires understanding multiple interconnected systems:
PAI_DIR,PAI_HOME,DA, andTIME_ZONEmust all be correctly configuredThe README alone spans over 4,000 lines. The interactive wizard added in v2.0 helps but introduces its own issues—Issues #259-262 document problems with unclear next steps, secrets coordination failures, and unclear pack contents before installation.
Recurring patterns of unaddressed issues
Several installation problems have persisted across versions without resolution:
The v2.0 transition was explicitly designed to address the "Jenga tower" problem, yet issues filed after the transition indicate core problems persist. Issue #230 reports that the history system isn't capturing data as documented. Issue #228 documents hook output format mismatches. The architectural improvements haven't eliminated the installation friction.
Maintainer responsiveness and community dynamics
Daniel Miessler remains active in GitHub Discussions and has acknowledged problems publicly. However, most individual issues lack direct maintainer responses. The community has stepped in—contributors like @hjbrandt, @noxx, @mellanon, and @smolcompute have been filing detailed bug reports and providing workarounds—but many critical issues remain open without official fixes.
The project's response to complexity has been to add more complexity: the v2.0 pack system, the interactive wizard, AI-assisted installation, extensive documentation. Each layer addresses symptoms while the fundamental problem—a system designed for one user's macOS environment that doesn't translate reliably to other contexts—remains.
Conclusion: architectural ambition exceeds installation reliability
The evidence suggests that PAI's installation difficulties stem from a core tension: the system was built as a sophisticated personal AI infrastructure that worked for its creator but wasn't engineered for reliable external deployment. The repeated breaking changes, the explicit acknowledgment that v1.x "didn't work," the unresolved Linux and Windows issues, and the pivot to recommending AI-assisted installation all point to a project where complexity has outpaced usability.
For users evaluating PAI, the documented pattern is clear: expect significant troubleshooting, particularly on non-macOS systems; be prepared for breaking changes between versions; and understand that the "AI-first" installation approach exists because the manual process has proven unreliable. The project's ambition is admirable, but the installation experience remains a significant barrier to adoption that the maintainers have acknowledged but not yet solved.
Beta Was this translation helpful? Give feedback.
All reactions