Skip to content

Generative Chatbots

What are we trying to accomplish?

This module takes students from understanding LLMs conceptually to deploying them as autonomous coding agents. The arc is deliberate: students first learn how large language models work and how to communicate with them effectively, then move to integrating an LLM into a real Python application via API, then shift to using a coding agent — Claude Code — as a development tool on a real full-stack project. Each sub-module answers a different version of the same question: what do you control, and how? In the first sub-module the answer is prompts. In the second, code. In the third and fourth, the agent's context, tools, rules, and lifecycle hooks. By the end, students can configure and direct a coding agent to read unfamiliar codebases, identify bugs, and execute multi-file changes — with the precision and intentionality of a senior developer, not a user clicking "accept suggestion."

Lessons

  1. Working with LLMs
  2. Integrating LLMs
  3. Coding Assistants
  4. Claude in Action

Module Topics

  • Transformer architecture and next-token prediction
  • Prompt engineering: CREF framework, zero-shot, few-shot, chain-of-thought, role prompting
  • Comparing LLM providers across context window, cost, multimodal support, and instruction-following
  • Gemini API authentication and model selection
  • Structured outputs with Pydantic schemas
  • Multi-turn conversation management and message history
  • Multimodal inputs and thinking mode via the Gemini API
  • Coding agent architecture: agent loop, MCP, tools, and LLM separation
  • Model Context Protocol (MCP) as a standard for tool integration
  • Claude Code installation, Ollama local model fallback, and token cost management
  • Context window management: @-path targeting, planning mode, thinking modes, interrupt, rewind, compact, clear
  • Playwright MCP for browser-based UI and security testing
  • Custom slash commands in .claude/commands/
  • PreToolUse and PostToolUse hooks for automated enforcement and review
  • Anthropic SDK integration inside hooks for programmatic LLM calls