โป vision
A personal AI infrastructure framework that turns Claude Code into a daily-driver exoskeleton. Multi-mode (NATIVE / ALGORITHM / MINIMAL), voice-announced phase transitions, a self-modifying agent that announces itself with my own cloned voice, and a strict ISC-criteria protocol for hill-climbing on every task. Currently version 4.0.3.
โป why
Default UX = default thinking. The shipped chat-and-tool-use UX of every AI lab is too generic for the way I actually work. PAI is what happens when I stop accepting a stranger's mental model for my own brain.
โป the algorithm
OBSERVE โ THINK โ PLAN โ BUILD โ EXECUTE โ VERIFY โ LEARN. Each phase has a voice announcement, an ISC checklist, a budget, and a written PRD. The agent edits the PRD as it works.
what's currently wired in
- โถvoice announcements (ElevenLabs cloned voice) at every phase boundary
- โถPRD-as-source-of-truth โ every task has a `MEMORY/WORK/{slug}/PRD.md`
- โถISC count gate โ refuses to exit OBSERVE without enough atomic criteria
- โถa custom statusline showing model, budget, and mode
- โถstructured reflection logs feeding a self-improvement skill
- โถthe entire framework is invoked by `claude` in any project โ no per-project setup
โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ โ PAI ยท the seven-phase ritual โ โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโค โ ๐ OBSERVE โ reverse engineer โ โ ๐ง THINK โ premortem, risks โ โ ๐ PLAN โ prereq + capability sel โ โ ๐จ BUILD โ invoke skills/agents โ โ โก EXECUTE โ do the work + tick ISC โ โ โ VERIFY โ test every criterion โ โ ๐ LEARN โ reflection JSONL log โ โโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโโ