Some questions and feedback mid-migration to v5.0

[stephtara]

Context: non-dev, I don't live in the terminal but I'm happy to play in it every now and then; some general programming knowledge but not enough to write a script on my own. Started with PAI 2 weeks ago, been trying to twist AI's arm to do my bidding for a while now, but no priori Claude Code experience.

I have a few general questions on how people are actually using PAI, predating v5.0, and a couple of remarks on the migration progress.

Q1 Languages
Is anybody successfully using their DA in two different languages? I'm a French-English bilingual and I sometimes switch depending on context. My DA started sticking "bilingual word salad" in its responses and documents at some point. I'd like it to follow whatever language I'm using. I'm also wondering if it makes sense to "force" English (the main language I tend to be using with it) in any kind of system files (memories, skills, Telos, etc). Some of my projects have subject matter that is actually in French (e.g. managing a francophone online community), others are simply topics I'm more fluent in in French (e.g. plant care). Interested in hearing about experiences and advices on that matter.

Q2 AI provider choice
I get that the idea of PAI is to be an abstraction layer above the LLM, so switching is possible. I've seen there is a port to OpenCode. But how close are we to "AI provider migration" that could be done without having to significantly rewrite portions of PAI? I'm trying to get a sense of what it would mean if I wanted, for example, to switch to ChatGPT. Is this something that could be doable by a tinkerer like me and their DA, or "fugeddaboudit"?

Q3 Claude model
With v4.0.3 I rather quickly found that using Opus as my base model was infinitely less frustrating than Sonnet (but oh, the token use, specially with my Pro plan). A lot of what I want my DA to do is relieving me of cognitive load (ADHD and post-concussion syndrome are a great combination). I'm not really doing any development at the moment – it's more keeping track of my umpteen ongoing projects and open loops, where was I at last time I worked on this, go through 26 years of blog posts and extract useful information about xyz, help me figure out what to eat tonight, that kind of stuff. My impression is that whatever I spare token-wise using Sonnet, I end up paying back in having to do more work myself to keep things on track or get what I actually wanted. What are your experiences with that? Is anybody here actually using Sonnet? For what kind of tasks?

Q4 Skills and "using PAI"
I'm struggling a little with the "actually using PAI" when one is starting out. I tend to try and interact according to the "assistant" metaphor but it seems to be hit-and-miss and I'm wondering if I shouldn't be directly invoking skills (or whatever the slash-commands are) rather than expecting the DA to know when to use what. How much do you type slash-things versus just "talk"? And did that evolve over time with the maturity of your DA personalisation? I also found myself rather quickly bogged down in a lot of "fixing the system" sub-projects: in particular, things around token use and memory management and task continuity.

• warning me when context got large enough that it would make sense to start a new session rather than continuing in the same one
• making sure all the important information from a conversation was captured when I signed off
• picking a path which offered a good balance between "token-economical" and "brain-economical"
• decent handoff between one session and the next (FWIW saw this fail again in v5.0: did a first part of the migration, ended session, DA in the loop about that and that we would resume, and next session when I talked about continuing migration, DA had no clue what I was talking about)

Q5 Oversight
How much do you "correct and control" what your DA writes in skills, Telos files, etc? Do you really just let it manage its thing, or do you intervene to fix things manually?

V5.0 Feedback
A few notes so far:

• loved the GUI for the installer, now waiting for one that will make the terminal conversations look more like a "chatting with AI" chatbot
• migration guide says to do interview before migration, which doesn't seem very logical, shouldn't it be the opposite?
• the command given at the end of the installer to launch PAI disappeared before I had a chance to copy it (not an issue for me, but could be for others)
• there is a lot of documentation about what PAI is and how it functions and is built/designed, but I haven't seen much about "how to actually use it"; at this stage, what I'm experiencing is that it's not yet "plug-and-play ready" (not criticising, it's an observation) and that "just try and use it" is a little frustrating (mind-blowing too of course).

Thanks for the vision, commitment and hard work that make PAI possible. I'm really very enthusiastic about it. And I perfectly understand that even moving fast, not everything can be done at the same time – that's OK. Just trying to get my bearings a bit more clearly.

Looking forward to reading about how things are for you on these topics!

[danielmiessler]

Hey @stephtara, thanks for this. Looks like it's been answered/resolved, and we're tidying up the tracker ahead of a big release where LifeOS moves to a skill-based install. Closing for now; reopen anytime if you're still stuck. Appreciate you.