01/ 06

Chapter One

I

The AI System Stack

Reference glossary

AI Concepts & Terminology

The AI landscape is full of terminology: LLMs, agents, RAG, context windows, knowledge bases, memory systems, harnesses, guardrails... Each term sounds important— but how do they fit together? Where does each concept live?

+

Source: Stanford HAI AI Definitions

System behavior

How the concepts lives

Open this worked example to see how the same concepts move from user dialogue into layered system behavior and a supervised harness.

+

Example - Callout Design Task: 6-Step Harness Flow

User Question (Natural Language):
  • "Build callouts style panel to add related items that matches the theme of AI Glossary card and handles errors gracefully. I need it production-ready."
Attachment: AI_GLOSSARY_TECHSHARE.md
Prompt
Memory
Harness Plans Steps & Applies Constraints:

Step breakdown:

  • Inspect current implementation
  • Implement mapping + callout behavior
  • Iterate from feedback and validate
Planning
Orchestration
Context Assembled (Conversation + Retrieved Code):

Harness loads:

  • Conversation context
  • Relevant files + data
  • Available tools
Memory
Knowledge
Context
Model Reasons & Proposes Changes:

LLM decides:

  • Select minimal change points
  • Prefer in-place updates
  • Plan interaction/state and trigger validation
Agent Loop - Decide
Prompt Eng
Tools Read/Edit/Validate:

Harness executes tool calls:

  • Read + edit targeted files
  • Run diagnostics and build
  • Loop results back for iteration
Agent Loop - Act/Obs
Function Calling
Skills
Key Gate
Safety Gate Checks Output Before Final Response:

Harness validates output criteria:

  • No file errors + build passes
  • Behavior matches request
  • Fail = iterate, pass = deliver
Evaluation
Validation
Human Approval
Outcome
Final Response to User:
  • Updated editorial.tsx
  • Implemented callout interactions + style updates
  • Updated part-1.tsx
  • Wired glossary cards to related data
  • Created part-1-techshare-crosswalk.ts
  • Added glossary-to-techshare mapping
  • Validated before delivery
Verification
Handoff

System map

The six-layer stack

Modern AI system has six layers. Previous AI products mostly lived at layer 2 + layer 6: a model plus a chat UI. Today's products span all six, and the engineering work that differentiates teams happens in the middle.

+
L6
Application

The product surface.

Chat, IDE, Canvas, Browser Extension, etc.

L5
Orchestration Layer

The harness that governs the model.

+Show L5 sublayers
L5.1 Specification & Memory
What the system stores from the directive
L5.2 Constraint Architecture
What rules are enforced in real-time
L5.3 Feedback Loop
What validations happen before output
L5.4 Human Oversight Gate
Where humans authorize before action
L4
Tools

Function calling, MCP, browser use, computer use.

The model gains hands.

L3
Context

What the model sees right now.

Knowledge + Memory + System Prompt + Tools + User's Latest Message

L2
Model

A function from tokens to tokens.

Claude, GPT, Gemini, Llama.

A component, not a product.

L1
Foundation

Inference servers, execution environment

Live fully. Be bold. Grow always.

Personal Knowledge Base

Quick reference for turning scattered source material into an organized, queryable knowledge base.

+

Three workflows

A - ingestObsidian

Add a new source

B - queryClaude Cowork

Ask a question

C - healthClaude Cowork

Lint the wiki

Vault map - 'Better Bold Brain'

My_Knowledgebase/
│
├── CLAUDE.md                  Standing instructions for the AI maintainer.
├── WORKFLOW.md                This file — the human's quick reference.
│
├── raw/                       SOURCE MATERIAL. Immutable. Read-only after ingest.
│   ├── README.md              raw/ folder rules and naming conventions.
│   ├── _TEMPLATE.meta.md      Sidecar template for non-markdown sources.
│   │
│   ├── _inbox/                ✏️  ONLY mutable folder. New files land here first.
│   ├── papers/                Academic / research PDFs.
│   ├── articles/              Blog posts, news, long-form web content.
│   ├── docs/                  Vendor docs, manuals, whitepapers.
│   ├── clippings/             Short web snips, tweets, threads.
│   ├── data/                  CSV, XLSX, JSON datasets.
│   ├── transcripts/           Whisper transcripts of audio/video files.
│   ├── assets/                Images embedded in clipped articles.
│   └── media/
│       ├── images/            Standalone images (diagrams, screenshots).
│       ├── audio/             Podcasts, interviews, voice notes.
│       └── video/             Talks, demos, recordings.
│
└── wiki/                      SYNTHESIZED KNOWLEDGE. The AI maintains this.
    ├── index.md               Catalog of every wiki page. Always current.
    ├── log.md                 Append-only session log of ingests/queries/lints.
    ├── overview.md            Evolving synthesis of the whole AI landscape.
    │
    ├── tools/                 One page per AI tool or framework.
    ├── models/                One page per LLM or embedding model.
    ├── concepts/              Durable ideas: RAG, evals, fine-tuning, agents.
    └── synthesis/             Cross-cutting analysis, comparisons, decisions.

Status legend

seedling
growing
evergreen
Promote when a page has 2+ sources and cross-links.

Idea inspiration

Andrej Karpathy - personal knowledge base setup

gist.github.com/karpathy/442a6bf5...

View gist

You Can Outsource Thinking, But Not Understanding.

Andrej Karpathy-- Andrej Karpathy