AI Tutorial - Claude Code & Agent Frameworks

ai
claude code
Author

Lawrence Wu

Published

March 10, 2026

Below is an AI Tutorial I did for some friends. Unfortunately my OBS resolution was set too high so not every frame was captured in the video.

Below are the notes used.

Intro

  • Why I’m doing this

Foundational Material

Language Models

Large Language Models Explained Briefly https://www.youtube.com/watch?v=LPZh9BOjkQs&t=1s

Deep Dive into LLMs like ChatGPT (Karpathy) https://www.youtube.com/watch?v=7xTGNNLPyMI&t=8425s

Step 1: Pretraining

Pretraining

Step 2: Tokenization

Tokenization

https://tiktokenizer.vercel.app/

Step 3: Neural network I/O

Neural network I/O

Neural network internals

Neural network internals

Show https://bbycroft.net/llm

  • Artificial Intelligence
    • Machine Learning
      • Supervised Learning - I have labeled data, emails of spam and not spam.
        • Label or not
        • Neural Networks
      • Unsupervised Learning - Clustering - customer data, I want to cluster my customer segments
    • Generative AI
      • Pretraining - thousands of terabytes of data
      • Post-training
    • 2023-2024

Inference

Inference

“Psychology” of a base model

“Psychology” of a base model

Post-training

Post-training

Agent Frameworks

https://lawwu.github.io/posts/2026-03-04-agentic-knowledge-work/

  • No LLM - pre 2023
  • LLM, Code Completion - 2023-2024
    • Simple LLM Chains
  • Coding IDEs - Cursor, Windsurf - 2024
  • Coding Agents - 2025

Third Era of Software Development

Third Era of Software Development

Thirty-five percent of the PRs we merge internally at Cursor are now created by agents operating autonomously in cloud VMs. We see the developers adopting this new way of working as characterized by three traits:

  1. Agents write almost 100% of their code.
  2. They spend their time breaking down problems, reviewing artifacts / code, and giving feedback.
  3. They spin up multiple agents simultaneously instead of handholding one to completion.

V1.0 Agent Frameworks

LLM + a library like LangGraph

V1.0 Agent Frameworks

V2.0 Agent Frameworks

LLM + a CLI like Claude Code

  • the underlying models have improved dramatically and now have reasoning capabilities
    • Chain of Thought - My prompt here - think step by step
  • the base tools that Claude Code has access to are very generic. For example with the Bash tool, the agent literally has access to thousands of battle-tested programs (CLIs).
  • The biggest change though has been code generation and the ability to create tools on the fly that the agent needs to solve a task.
  • You can run these agents in parallel

V2.0 Agent Frameworks

This will continue to improve as

  • models improve
    • there will be different types of models too - world models
  • harnesses improve

Claude Code

Bash Basics

Basics

Try it out (10 mins)

On a problem you want to solve.

In ai-tutorial VS Code terminal:

i have downloaded cross-references from openbible, can you build a visualization of this data for me?

Can show the process:

  • don’t need to specify the path of the data
    • agentically finds the data
  • does some data verification without me asking
    • verifies all 66 books are present
  • decides what visualization to build “interactive chord diagram”
❯ i have downloaded cross-references from openbible, can you build a visualization of this data for me?

⏺ Let me first look at the data to understand its structure.

⏺ Read 3 files (ctrl+o to expand)

⏺ 344,800 cross-reference rows. Let me examine the data more and then build a visualization.

⏺ Read 2 files (ctrl+o to expand)

⏺ Good — all 66 books present. Let me build an interactive chord diagram visualization.

Skills and Plugins

Plugins - groups of skills

Try it out (10 mins)

Install a plugin and use a skill from a third-party.

Harness Engineering

Chaining together skills

In agentic-ml-plugin VS Code terminal:

claude --plugin-dir ./plugins/agentic-ml --dangerously-skip-permissions

Agentic Knowledge Work

Some action items:

  • meta-work - ask, is there a better way to be doing this? an agent-first way?
  • learn the tools
    • it takes time to learn how to use a new tool. Claude Code and Codex both have different functionality. The underlying models are slightly different. 80% of it is the same though.
    • AGENTS.md / CLAUDE.md
    • Have your agent use CLIs. Create CLIs.
    • meta
      • Claude Code can teach you about itself
      • Claude Code can configure itself - “configure a statusline that shows total token cost for the session”
  • be constantly learning
  • Don’t outpace your understanding: Review plans, read diffs, and make sure you understand the system well enough to validate what the agent changed
  • Still think for yourself

OpenClaw

Always-on agents