Skip to content

jwm-axoni/auggie-pai

BranchesTags

Repository files navigation

Auggie PAI

Personal AI Infrastructure for Augment Code CLI

Algorithm • Architecture • Skills • Installation • Customization

What is Auggie PAI?

An adaptation of Daniel Miessler's PAI (Personal AI Infrastructure) for the Augment Code CLI.

PAI transforms AI coding assistants from stateless chat into a structured execution system with persistent memory, verifiable criteria, and repeatable methodology. Auggie PAI brings that same rigor to the Augment CLI environment.

What you get:

  • A 7-phase Algorithm that structures every non-trivial task
  • 18 skills across thinking, research, security, and personas
  • Persistent memory that learns from every session
  • ISC (Ideal State Criteria) — atomic, binary-testable criteria that define "done"
  • PRD tracking — structured work artifacts that survive context loss

The Algorithm

The Algorithm — 7-Phase Execution Loop

The Algorithm is PAI's core execution engine. Every non-trivial request flows through 7 phases:

Phase What Happens
OBSERVE Reverse-engineer the request. Extract explicit wants, implied wants, what's NOT wanted. Generate ISC criteria. Select capabilities.
THINK Pressure test. Premortem. Identify riskiest assumptions. Refine criteria through the Splitting Test.
PLAN Validate prerequisites. Create execution plan. File-edit manifest for multi-file work.
BUILD Create artifacts. Invoke each selected capability. Constraint checkpoint after each artifact.
EXECUTE Run the work. Continuously verify against criteria — don't batch to end. Mark PRD checkboxes as criteria pass.
VERIFY Mechanical verification of every criterion. No rubber-stamping. Specific evidence required.
LEARN Reflect. Capture patterns. Write to persistent memory for future sessions.

The Learning Loop — How Auggie Gets Smarter

Mode Selection

Not everything needs the full Algorithm:

Mode When Example
MINIMAL Greetings, acknowledgments "hello", "thanks"
NATIVE Quick tasks under 2 minutes "what's 2+2", "rename this var"
ALGORITHM Everything non-trivial "build a caching layer", "debug this auth flow"

Effort Levels

The Algorithm scales to match the task:

Tier Time Budget ISC Criteria Min Skills
Standard <2 min 8–16 1–2
Extended <8 min 16–32 3–5
Advanced <16 min 24–48 4–7
Deep <32 min 40–80 6–10
Comprehensive <2 hrs 64–150 8–15

ISC — Ideal State Criteria

The secret sauce. Every task is decomposed into atomic, binary-testable criteria before any work begins.

Each criterion must pass the Splitting Test:

  1. "And"/"With" test — joining two verifiable things? Split.
  2. Independent failure test — can part A pass while B fails? Split.
  3. Scope word test — "all", "every", "complete"? Enumerate what "all" means.
  4. Domain boundary test — crosses UI/API/data boundaries? One criterion per boundary.
Bad:  "Blog publishing handles draft to published with SEO metadata"
Good: "Draft status stored in frontmatter YAML field"
      "Published timestamp set on first publish only"
      "Meta description under 160 characters"
      "Canonical URL set to published permalink"

System Architecture

System Architecture — Core Components

Five Core Components

Algorithm (rules/algorithm.md) — The 7-phase execution loop. Injected into every prompt via alwaysApply: true frontmatter. This is the brain.

Skills (skills/) — 18 self-contained methodologies the Algorithm can invoke during BUILD/EXECUTE. Each has a SKILL.md defining triggers, methodology, and output format.

Memory (MEMORY/) — Persistent across sessions. Reflections from LEARN phases, user corrections, synthesized patterns, and active work state.

Memory System — Persistent Across Sessions

PRD Tracking (.prd/) — Product Requirements Documents created during ALGORITHM mode. YAML frontmatter tracks phase, progress, effort. ISC criteria live as checkboxes. Survives context loss.

Security (rules/security.md) — Prompt injection defense, credential handling, SSRF awareness, destructive operation guards.

Skills

Thinking

Skill What It Does
first-principles Deconstruct → Challenge assumptions → Reconstruct from ground truth
council Multi-perspective debate with 3 rounds of deliberation (see below)
red-team Adversarial analysis — steelman then attack
iterative-depth Multi-angle exploration through cognitive lenses
science Hypothesis-driven investigation with experiment design

The Council — Multi-Perspective Debate

Research

Skill What It Does
quick-research Rapid research pass
standard-research In-depth multi-source research
deep-investigation Comprehensive investigation
extract-knowledge Structured knowledge extraction
extract-wisdom Wisdom and insight extraction

Security

Skill What It Does
recon Network and domain reconnaissance
web-assessment OWASP-style web application security
threat-model Threat modeling and risk analysis
prompt-injection LLM prompt injection testing

Agent Personas

Persona Identity Specialty
engineer Marcus Webb Implementation, debugging, optimization
architect Serena Blackwood System design, patterns, trade-offs
qa-tester Quinn Torres Testing, edge cases, quality gates
pentester Rook Blackburn Offensive security, vulnerability assessment

Slash Commands

/think <skill> <topic>      Apply a thinking methodology
/research <mode> <topic>     Run research (quick/standard/deep/extract)
/assess <type> <target>      Security assessment
/agent <persona> <task>      Adopt an agent persona
/memory <operation>          Memory operations (read/status/reflect)
/status                      Show PAI system status

Installation

Fresh Install (No Existing Auggie Config)

git clone https://github.com/jwm-axoni/auggie-pai.git ~/.augment
cd ~/.augment && ./setup.sh

The setup script creates all directories, initializes memory files, and copies settings.json.example to settings.json. Edit it with your credentials and paths.

Existing Auggie Users

Already have Auggie CLI configured with your own settings.json, MCP servers, and sessions? This merges PAI into your existing setup without touching any of that.

# Back up first
cp -r ~/.augment ~/.augment.backup

# Clone to a temp location
git clone https://github.com/jwm-axoni/auggie-pai.git /tmp/auggie-pai

# Copy PAI components into your existing config
cp -r /tmp/auggie-pai/rules ~/.augment/
cp -r /tmp/auggie-pai/skills ~/.augment/
cp -r /tmp/auggie-pai/commands ~/.augment/
cp -r /tmp/auggie-pai/USER ~/.augment/
cp -r /tmp/auggie-pai/.prd ~/.augment/
cp /tmp/auggie-pai/CLAUDE.md.example ~/.augment/
cp /tmp/auggie-pai/settings.json.example ~/.augment/

# Run setup to create memory directories
cd ~/.augment && bash /tmp/auggie-pai/setup.sh

# Clean up
rm -rf /tmp/auggie-pai

Then add these hooks to your existing settings.json:

"hooks": {
  "SessionStart": [{
    "hooks": [{
      "type": "command",
      "command": "cat $HOME/.augment/MEMORY/STATE/current-work.json"
    }]
  }],
  "Stop": [{
    "hooks": [{
      "type": "command",
      "command": "echo '{\"reminder\": \"check_reflection\"}' >> $HOME/.augment/MEMORY/STATE/stop-reminders.log 2>/dev/null || true"
    }]
  }]
}

If you already have a hooks key, merge these into your existing hooks object.

Configure

File Purpose
settings.json MCP servers, credentials, indexing paths
USER/ABOUTME.md Your role, expertise, timezone
USER/AISTEERINGRULES.md AI behavior preferences
USER/WORK/ Company context and active projects

Verify It Works

Launch Augment CLI and type: analyze the pros and cons of Redis vs Memcached

You should see the Algorithm activate with:

  • ♻︎ PAI ALGORITHM (v3.0-auggie) header
  • Reverse engineering output
  • ISC criteria generation
  • All 7 phases executed

Directory Structure

~/.augment/
├── rules/                       # Always-injected rules
│   ├── algorithm.md             # The Algorithm v3.0-auggie
│   ├── security.md              # Security guidelines
│   └── context-routing.md       # Topic → file routing
│
├── skills/                      # 18 invocable skills
│   ├── first-principles/        # Thinking skills
│   ├── council/
│   ├── red-team/
│   ├── engineer/                # Agent personas
│   ├── architect/
│   ├── quick-research/          # Research skills
│   ├── recon/                   # Security skills
│   └── ...
│
├── commands/                    # /slash command definitions
│
├── USER/                        # Your configuration
│   ├── ABOUTME.md
│   ├── AISTEERINGRULES.md
│   └── WORK/
│
├── MEMORY/                      # Persistent memory
│   ├── LEARNING/                # Reflections, corrections, patterns
│   ├── STATE/                   # Active work pointer
│   └── RESEARCH/                # Research archives
│
└── .prd/                        # PRD work tracking
    └── templates/

Customization

Add a Skill

mkdir -p ~/.augment/skills/my-skill

Create SKILL.md:

---
name: my-skill
description: Short description. USE WHEN trigger words go here.
---

# My Skill

## Methodology
1. Step one
2. Step two
3. Step three

Add a Slash Command

Create commands/my-command.md:

---
description: What this command does
argument-hint: "<arg1> <arg2>"
---

Instructions for the AI when this command is invoked.

Tune AI Behavior

Edit USER/AISTEERINGRULES.md to adjust response style, technical preferences, and operational boundaries.

How It Compares

Feature Vanilla Augment CLI With Auggie PAI
Task structure Freeform 7-phase Algorithm with ISC
Memory None between sessions Reflections, patterns, corrections
Skills Generic capabilities 18 specialized methodologies
Work tracking None PRD with YAML frontmatter + criteria checkboxes
Security Basic Prompt injection defense, credential guards, SSRF awareness
Quality gates None Mechanical verification, anti-criteria, Splitting Test

Credits

AI should magnify everyone — not just the top 1%.

About

PAI (Personal AI Infrastructure) v3.0 adapted for Augment Code CLI — structured 7-phase Algorithm, 18 skills, persistent memory, ISC methodology

Topics

cli pai personal-ai ai-infrastructure augment-code

Resources

Readme
Activity

Stars

3 stars

Watchers

0 watching

Forks

0 forks
Report repository

Releases 1

PAI v3.0-auggie — Initial Public Release Latest
Mar 4, 2026

Packages

 
 
 

Contributors

Languages