Skip to main content
Version: 25fa

Coding with AI

Brief intro to generative AI​

https://docs.google.com/presentation/d/1Z5f5ntyQ6xhSC160Go66ej8Kkh3UmkheGyYoOfZRk2k/edit?usp=sharing

The rest of this lecture is summarized from research literature, blogs, and personal accounts.

Some papers:

Where do programmers code with AI?​

  • The web interface is good for generating a whole thing from start to finish
    • Less interactive experience
  • IDE with inline suggestions (extreme auto-complete) and chat interface is good for small parts of a large thing
    • More interactive experience
  • Claude Code is a command-line tool that runs from the terminal and executes multi-step tasks
    • Can range from very interactive to completely autnomous

Common concerns​

  • Concerns about over-reliance (self-reported by students)
  • Hallucinations and incorrect information
  • Security flaws
  • Outdated information
  • Access inequality (paid vs. free versions)
  • Reduced deep comprehension for complex tasks
  • Some studies show AI slowa down developers (https://arxiv.org/pdf/2507.09089)

Each task in the PDI process can use AI in some way:​

Systematic Software Design Process

AI can help us do these efficiently, but it cannot replace human creativity and understanding of the codebase.​

  1. Identify what domain-related information the AI needs
  2. Engage with the AI through prompts that state the need as a structured request with specific information, context, the problem statement, and desired outcomes
  3. Evaluate the AI output against domain knowledge, expected results, and other success criteria
    1. Often run and observe in a sandbox
  4. Calibrate by providing the AI with feedback and additional context
  5. Tweak the AI-generated artifacts to better fit standards
  6. Finalize and write documentation including decisions and rationale

This workflow can be used for any of the tasks in the PDI process.

Why we discourage "vibe coding" in the PDI sequence:​

  • "Vibe coding" is where you only evaluate the execution of the AI-generated code, and not the code itself
  • Vibe coding leads to "productivity collapse"
  • When we write code, we learn about how our codebase works (and about how to write better code)
  • If we don't understand our own code, it's very hard to troubleshoot, give the right feedback to the AI, add to the codebase, etc.
  • So we have the AI write its own brittle fixes
  • This tends to spiral into branching dependencies where nobody understands how it works

Common uses for the web interface​

  • Learning new syntax (for example, when using a new package)
  • Explaining error messages to help debug
  • Generating code
    • Generate initial code for a starting point, which you refine there or in an IDE
  • Reviewing code
    • Find bugs we might have missed
    • Think of test cases we didn't consider
    • Make things more efficient / easier to read
  • Translating between programming languages

Common uses for Cursor / Windsurf / Github Copilot IDE plugin​

  • Generate boilerplace code quickly (class with constructor, test class with setup, if __name__ == '__main__', etc)
    • Especially if you already have examples using of how you prefer it in your codebase
  • Figure out bugs that require understanding interaction between multiple files (using the IDE's chat feature)
  • Make modifications that require modifying parts of multiple files
    • Context is any files you have open in the editor
  • Suggest commands that you can choose to run in the IDE's built-in terminal

Common uses for Claude Code in the terminal​

  • Autonomously run commands in the terminal (installing stuff, running tests)
  • Do multiple things at once (e.g., run commands and write code)
  • Automate repetitive coding tasks
  • Do Git stuff (commits, PRs, etc.)
  • Refactor code accross entire codebases
  • Write code that requires understanding dependencies / how many files interact
  • Automate entire workflows like test-driven development
  • Work for hours without human interaction (this scares Rasika but it is a common use)
  • Split tasks and delegate specialized tasks to sub-agents
  • Can run as a script in headless mode (no need for a terminal)

Note: There's also a Claude Code VSCode extension (beta) that provides a graphical interface for Claude Code within VSCode, offering a middle ground with inline diff previews and one-click rollback while maintaining Claude Code's agentic capabilities.