PAI: Turn Claude Code Into a Life Operating System

PAI (Personal AI Infrastructure) is an open-source system that turns Claude Code into a life operating system. Not just a coding assistant. A single AI interface for your goals, your work, your projects, your patterns. Built by Daniel Miessler, the creator of Fabric. 12k+ stars on GitHub.

It runs natively inside Claude Code. No wrappers, no separate app, no vendor lock-in. One install command copies a structured set of skills, workflows, hooks, and memory templates into your ~/.claude/directory. From there, everything Claude does is filtered through who you are and what you're building.

The problem nobody talks about

Claude Code is powerful out of the box. But it forgets you between sessions. It doesn't know your goals, your preferences, or what you were working on last week. Every conversation starts cold.

Most people solve this with a CLAUDE.md file. Some add memory tools or custom instructions. PAI takes that idea and builds an entire operating system around it: structured memory, a goal-tracking system, a local dashboard, and 45 skills that handle everything from writing to research to task decomposition.

The pitch: stop configuring Claude from scratch every session. Define yourself once, and let the system carry that context everywhere.

Why this changes everything

Four things set PAI apart from a CLAUDE.md file and some custom prompts.

Telos

After install, PAI runs you through a /interview command. Four phases: mission, goals, beliefs, wisdom, challenges, mental models. The system calls this your Telos. It's stored in PAI/USER/TELOS/ and every skill, workflow, and decision Claude makes references it. Without Telos, the system has nothing to optimize against. With it, Claude stops being a generic assistant and starts operating in your context.

The Algorithm (v6.3.0)

PAI doesn't just respond to prompts. It runs a seven-phase loop modeled on scientific methodology: OBSERVE, THINK, PLAN, BUILD, EXECUTE, VERIFY, LEARN. A mode classifier evaluates each prompt and picks a complexity tier (E1 through E5). Simple questions get minimal processing. Complex projects get the full loop with verification checkpoints and advisor calls at commitment boundaries.

45 skills, 171 workflows, 37 hooks

Not prompt templates. Structured task sequences that run inside Claude Code natively. Skills are organized across 12 categories. Hooks fire at eight event points: SessionStart, UserPromptSubmit, PreToolUse, PostToolUse, Stop, SubagentStop, PreCompact, SessionEnd. The routing goes Code, then CLI, then Workflows, then SKILL.md. Prompts wrap code. Code doesn't wrap prompts.

Pulse

A local dashboard running at localhost:31337. Tracks your goals, projects, and session patterns. 22 API routes. Optional voice notifications through ElevenLabs (falls back to desktop notifications without it). Health check, wiki access, and Telegram/iMessage bridges if you want them.

How PAI is built

The entire system is plain text. Markdown files and TypeScript. No database, no embeddings, no RAG pipeline. Search uses ripgrep. Miessler dropped RAG in June 2025 because filesystem search with ripgrep gave better fidelity with less complexity.

Memory is structured into six categories: WORK (active tasks), KNOWLEDGE (people, companies, ideas, research), LEARNING (meta-patterns), RELATIONSHIP (DA-to-user notes), OBSERVABILITY (tool calls, hook firings, satisfaction signals), and STATE (session registry).

Privacy is enforced at the architecture level. Every directory declares a containment zone. A ContainmentGuard hook blocks cross-zone data leaks before any tool use. 12 security gates run on every public release.

Step 1: prerequisites

You need three things installed:

• Bun (JavaScript runtime)
• Git
• Claude Code

Optional: an ElevenLabs API key for voice notifications. Without it, you get desktop notifications instead.

Step 2: install PAI

One command:

curl -sSL https://ourpai.ai/install.sh | bash

The installer backs up your existing ~/.claude/ to ~/.claude.backup-{TIMESTAMP}, verifies your prerequisites, prompts for the ElevenLabs key if you want voice, and copies the PAI structure into ~/.claude/.

If you prefer a manual install:

git clone https://github.com/danielmiessler/Personal_AI_Infrastructure.git
cd Personal_AI_Infrastructure/Releases/v5.0.0
cp -R .claude ~/
cd ~/.claude && ./install.sh

Step 3: run the interview

Start Claude Code and run:

/interview

This walks you through four phases:

1. TELOS captures your mission, goals, beliefs, wisdom, challenges, books, mental models, and narratives
2. IDEAL_STATE defines what success looks like
3. Preferences locks in your tools, conventions, working style
4. Identity tunes your Digital Assistant's personality

This is the most important step. Everything PAI does downstream depends on what you put in here. Take it seriously. The files land in PAI/USER/TELOS/ and PAI/USER/DA_IDENTITY.md.

Step 4: verify the install

Check that Pulse is running:

curl -s http://localhost:31337/api/pulse/health | jq

Open the dashboard in your browser:

http://localhost:31337

Test voice (if you set up ElevenLabs):

curl -s -X POST http://localhost:31337/notify \
  -H "Content-Type: application/json" \
  -d '{"message": "Hello from your DA"}'

Verify your DA identity is populated:

cat ~/.claude/PAI/USER/DA_IDENTITY.md

Step 5: use it

Start Claude Code in any project. PAI is now active globally. The Algorithm classifies each prompt automatically. Skills and workflows are available via slash commands. Memory persists across sessions.

Try asking Claude something that references your goals:

What should I work on today based on my current priorities?

If you ran the interview properly, Claude pulls from your Telos and gives you a prioritized answer. Not a generic productivity list. Your actual priorities.

Honest limitations

It overwrites your ~/.claude/ directory. The installer backs it up, but if you have an existing Claude Code setup with custom skills, hooks, and memory, you need to manually merge after install. PAI is opinionated about directory structure and expects its own layout.

Heavy.45 skills, 171 workflows, 37 hooks. That's a lot of context for Claude to load. If you use Claude Code as a quick bug-fix tool, this is overkill. PAI is built for people who want Claude as their primary interface, not a secondary tool.

Telos quality determines everything. Garbage in, garbage out. If you rush the interview or give vague answers, the entire system underperforms. The more specific and honest your Telos, the more useful the system becomes.

ElevenLabs dependency for voice. The voice features need a paid ElevenLabs account. Without it, you get desktop notifications, which work fine but lose the ambient awareness Miessler designed around.

Single-user. One DA per person. Not designed for team use or shared environments.

Quick reference