Daniel Miessler: Your AI Assistant Should Know Who You Are, Not Just What You Asked

TLDR:

Daniel Miessler demoed his completely rewritten Personal AI Infrastructure (PAI v2.4) at [un]prompted 2026. The system turns Claude Code into a persistent, goal-aware digital assistant through 7 architectural components, a scientific-method execution algorithm, and 67 domain-specific skills. The central argument: scaffolding matters more than model choice, and personal context transforms generic AI into something genuinely useful.

The Problem With Every AI Tool You're Using Right Now

Every conversation you start with an AI assistant begins at zero. It doesn't know who you are, what you're trying to accomplish, or what worked and failed last time. You explain yourself repeatedly, to a system that forgets you the moment the session ends. This is the problem Daniel Miessler came to [un]prompted 2026 to solve.

Miessler is the founder of Unsupervised Learning, a security researcher, and someone who has been building personal AI systems longer than most people have been seriously thinking about them. His project, PAI (Personal AI Infrastructure), is not a wrapper around a chatbot. It's a framework for building what he calls a Digital Assistant — a persistent system that loads your context, goals, and accumulated knowledge at the start of every session, runs tasks through a structured algorithm, and gets measurably better over time.

Version 2.4, which he presented at the conference, is a complete rewrite. Seven components, an updated algorithm (v0.2.23), a new memory system (v7.0), and 67 skills covering everything from blog publishing to security research. The whole thing runs natively on Claude Code, which Miessler chose specifically because its hook system, context management, and agentic architecture fit the requirements better than anything else available.

Seven Components, One Coherent System

The architecture is worth walking through because each component solves a real failure mode in how most people use AI today.

Intelligence is the combination of the model and the scaffolding it runs inside. Miessler's position here is direct: scaffolding beats raw model capability. A well-designed system running on a good model outperforms a poorly-designed system running on a great model. This is not a marketing claim, it's a design principle that shapes every other decision in PAI.

Context is everything the system knows about you across three tiers. Session memory covers the last 30 days of transcripts. Work memory holds structured project tracking. Learning memory captures accumulated patterns from interactions — what worked, what failed, and why. When you start a new session, the system is already oriented toward your current situation.

Personality gives the DA consistent traits, interaction style, and voice identity. Miessler's DA is named Kai, with a voice via ElevenLabs that stays consistent across tasks. This sounds cosmetic until you consider how much of effective collaboration depends on a shared working style. The alternative is an AI that's confidently helpful one moment and apologetically hedging the next.

Tools are structured in a hierarchy: code before CLI, CLI before prompts, prompts before skills. Deterministic solutions get tried first. Only when structured approaches won't work does the system reach for more flexible (and less predictable) methods. This reduces failure modes substantially.

Security gets its own component, which is not something most personal AI setups think about seriously. PAI implements policy-based permission validation before tool calls, constitutional principles embedded in the system, and behavioral guardrails. The goal is defense-in-depth that doesn't block normal workflows.

Orchestration covers 17 hooks firing across lifecycle events. When a task starts, context gets loaded. When it completes, outputs get captured. When errors occur, recovery procedures kick in. The system knows what phase it's in at all times.

Interface is CLI-first, with push notifications via ntfy or Discord and voice updates from ElevenLabs. The point isn't aesthetic. Terminal interfaces are faster, more scriptable, and more composable than GUIs. The system is designed for someone who actually works in a terminal.

The Algorithm: Scientific Method as Execution Loop

The part of Miessler's talk that resonates most is the algorithm at the center of PAI. Most agentic systems have some version of "plan, act, observe." PAI's algorithm is more structured, running as nested loops borrowed directly from the scientific method.

The outer loop tracks current state versus desired state. The inner loop runs seven phases: Observe, Think, Plan, Build, Execute, Verify, Learn. The thing that makes this concrete rather than aspirational is the Ideal State Criteria (ISC) — granular, binary, testable success conditions created during the Observe phase and checked during Verify. The system knows whether it succeeded, not just whether it produced output.

This matters because most agentic failures are not about the model making things up. They're about the system not knowing what "done" looks like. Miessler's algorithm is opinionated about this in a way that most frameworks aren't.

Skills and the 67-Pattern Library

PAI v2.4 ships with 67 skills and 333 workflows. A skill is a self-contained package combining domain knowledge, procedures, and CLI integrations for a specific area. Blog publishing, security analysis, research synthesis, code review — each is a distinct skill with its own context and tooling.

Personal skills (prefixed with underscore) contain API keys and user-specific data. System skills are shareable. When Claude Code starts, all skills load into the system prompt. Every request gets routed through the skill hierarchy automatically.

Miessler also integrates Fabric, a collection of 200-plus specialized prompts for specific tasks. The combination gives the system a wide coverage of common knowledge-work patterns without requiring custom prompts for every situation.

TELOS: Start With Your Goals

Before any of the technical components matter, PAI asks you to do something most people skip: define your purpose. The TELOS framework records your mission, top goals, biggest challenges, and active projects in structured files that load at the start of every session.

This is not a philosophical exercise. It's what makes the difference between an AI that answers questions and one that advances your actual objectives. When the system knows you're trying to build a newsletter business, your security research background, and that you're currently blocked on distribution, it generates responses calibrated to that context. Generic answers stop being useful.

The cardiologist example from the talk makes this concrete. A cardiologist who does bug bounty hunting as a side interest switched from standard Claude Code to PAI. His vulnerability discovery rate went up substantially because the system was trained on his specific techniques, loaded his domain context, and routed requests through skills designed for his workflow. The model didn't change. The scaffolding did.

What This Means in Practice

I keep thinking about the gap between how most developers use AI and how systems like PAI use it. The default approach is prompt-and-response. You ask, it answers, you copy the useful parts. PAI treats the AI as infrastructure — something that runs continuously, accumulates knowledge, gets better over time, and actually knows the difference between a great output and a mediocre one because it knows what you're trying to accomplish.

The open-source repository is live at github.com/danielmiessler/Personal_AI_Infrastructure. The architecture is designed so USER/ holds your customizations and SYSTEM/ holds the PAI infrastructure — you can upgrade the system without losing your data.

The 23-minute talk is worth your time if you're thinking seriously about personal productivity infrastructure. Not because you should clone Miessler's setup directly, but because the architectural decisions he's made are a useful model for thinking about your own.

Daniel Miessler - Anatomy of an Agentic Personal AI Infrastructure | [un]prompted 2026