(2026-09-11) Trying Kaminski's PKAI

There are lots of frameworks/etc to layer on top of AI tools like Claude Code for non-coding knowledge work/etc. I decided to try out Peter Kaminski's PKAI starter kit. (I know him, and he has a Discord support-community...)

Of course the first decisions are: where to put the "hq" folder, and what to name it (his example is johnsmith-hq). I decided to follow his naming convention, and put the folder inside my ~/Documents/fluxent folder, which is the level where I've been putting non-coding ClaudeCode stuff.

(Oooh, in launching Claude Code, it made me log in again, and in that process asked permission to connect to Claude chat sessions, which is something that's been bugging me. I also want to:)

• get voice working in various places (mainly phone+car)
• stop using Google Gemini, to focus context-building
• maybe "migrate" past Gemini sessions into Claude?

Name it Wilson (from CoachBot past). Pick Chief of Staff persona. Put it in ~/Documents/fluxent/agents/wilson/.

• it creates a bash-shortcut for wilson to launch ClaudeCode and move into that directory.

Setup finish: Setup's basically done. Two things before I hand you off:

• The geometry — Wilson lives in its home and always starts there; your HQ (billseitz-hq) is the hub — the switching center and growing knowledge base where smaller projects live in projects/. But it's not the only place Wilson works: bigger projects can live in their own folders anywhere, and Wilson is fully in bounds helping there too.
• The kit itself stays around as reference — inside it, pkai-starter/ has four mini-wikis (getting-started, git-guide, obsidian-reference, project-management) covering the way of working this kit assumes. Three ways to use them, not exclusive: Wilson can read them itself as background, we can walk through them together now or later, or just let bits surface naturally as topics come up. Want a walkthrough now, or leave it for later? (I chose to have Wilson read them itself for background.)
• And with that — setup's done. The next conversation belongs to Wilson, not me: open a fresh terminal, type wilson, and say hi.

Start a session

• it asks what I'm working on - I say "thinking about writing about Effectuation vs Coherence... but first, for context, read my existing *.md files up a level"
• this reminds me that I think I need to build out a little Python script to scrape a bunch of related-pages into a single big local Markdown file to feed Claude....

Sept15

• plan to join a Pete dyad-jam session today. Ask in Discord what prep is needed/useful.
• also feel like I just need to do some writing on this on my own, to start? (Examine the "conflict" under different contexts, rather than treating it as a single issue.)
  • actually, I should force myself to do this via the agent, to see how it feels
• also, blech I hate reading Markdown in the terminal, even 1 screen's worth, but esp if longer... there was talk of using Obsidian/MarkEdit....
  • duh, I'll ask wilson
• go to terminal - hmm that window is frozen, other terminal windows are fine. Close/restart.
  • "do you remember anything about Effection vs Coherence?" No. OK, we'll restart.
  • use prep for clear to save session-so-far. Also wilson asked if he should prompt me to do that periodically, and I said yes
• also see "update installed restart to update", so quit again, ask for version -> 2.1.272, which is current version. So start wilson again.
• launch Obsidian, tell it to open entire fluxent folder as a vault - maybe that's dumb since only a few folders are relevant... but sticking with that...
  • also told wilson to write long responses to existing outbox instead of terminal
  • (later during session) Pete notes benefit of using MeetingWords as reading-surface because human and agent can write simultaneously
• Pete shared a MD file about privacy thinking, I downloaded and copied into billseitz-hq, read in Obsidian.
  • it talks about buzz.xyz, and "relay" which is a buzz/nostr reference.
  • do I need a nostr keypair for the session? I asked.
  • answer: for now don't have to worry about buzz/nostr, most work will be in MeetingWords
    • Freya has an identity on the relay: a key, a membership row, a profile, a seat in the public channel. She has no harness, no daemon, and no subscription. When Pete wants her in the room, he commissions a background instance to visit. It is the same identity with the same charter, running as a disposable, scratch-only clone (see Foreground and Background). It reads the channel with the buzz command-line tool, takes part in channels Pete is in, writes a note on what it saw and did, and exits. That's the whole pattern. It's also how our agents already work on shared MeetingWords pads with other households' agents, so Buzz becomes a second venue for a habit we already had rather than a new kind of thing.
    • Three invitations arrived in one week, all shaped the same way: contribute pages to a collaborator's repository, join channels on our relay, take part as a peer in dyad jams on shared pads. Each time, the question was what may an agent of Pete's say, publish, or contribute in a place that isn't the estate, and how does it know without asking him every turn? Each time, the answer was a ruling in the moment, which is the wrong altitude.