LLM Coding Z Talk

Using LLMs to help when building software.

blackary.com/llm-coding-z-talk

LLM Coding Z Talk, 2025-08-13

What we'll cover

  • The Tooling Landscape Today
  • Where are AI Coding Tools Most Useful?
  • Obvious Risks and Pitfalls
  • LLM Workflows I use most
  • Tips and Tricks
  • Demo
  • Q&A
LLM Coding Z Talk, 2025-08-13

Caveats

  • This is largely based on personal reading and experience, more than deep understanding of LLMs
  • This is already out of date
LLM Coding Z Talk, 2025-08-13

The Tooling Landscape Today

Dedicated coding tools:
- Cursor
- Claude code
- (Aider, etc.)

LLM Coding Z Talk, 2025-08-13

The Tooling Landscape Today

Agents
- Codex
- Github Copilot
- Cursor Agents

(can check https://prarena.ai/ for one measure of performance):

LLM Coding Z Talk, 2025-08-13

The Tooling Landscape Today

Cloud-based coding tools:
- Replit
- Lovable
- Bolt

LLM Coding Z Talk, 2025-08-13

The Tooling Landscape Today

Standalone Chat tools:
- ChatGPT
- Perplexity
- etc.

LLM Coding Z Talk, 2025-08-13

The Tooling Landscape Today

Today's Top Models
- GPT-5
- Claude 4
- Gemini 2.5

(can check https://www.swebench.com/ for one measure of performance):

LLM Coding Z Talk, 2025-08-13

Where are AI Coding Tools Most Useful?

  • Repetitive transformations (refactoring, etc.)
  • Researching and applying code examples
  • Updates with feedback
    • e.g. "Fix then run the tests", "Fix then query the table", etc.
  • Popular packages
  • Typed code (static typed languages, python + type hints)
  • Understanding existing codebases
  • Contexts where you (or the reviewer) are able to judge quality and architectural decisions
LLM Coding Z Talk, 2025-08-13

Obvious Risks and Pitfalls

  • Looks good, must be right?
  • No need to learn new tools, just outsource them
  • Is this a good pattern?
  • "Code's compiling" (https://m.xkcd.com/303/)
  • Inefficiencies
    • Code way too complex/long/ugly
    • Code performance could be much better
LLM Coding Z Talk, 2025-08-13

LLM Workflows I use most

  • Claude for vibe-coding -- "give me a website that does x"
  • Cursor as extra-powerful auto-complete
  • Cursor as agent
    • Choose a model
    • Describe the task, how to test it, etc.
    • Let it go, review code/output, provide feedback, repeat
LLM Coding Z Talk, 2025-08-13

Tips and Tricks

  • Let the models fight
    • "My coworker wrote this, but I don't think it's right"
  • You're the PM
  • Ask it to make a plan for the approach it will take, write down steps, then review it
  • Learn to use the tool's pre-instructions (e.g. .cursor/rules, CLAUDE.md)
  • Give it the right context--point to files, docs urls, etc.
LLM Coding Z Talk, 2025-08-13

Tips and Tricks

  • Set up good rules in your repo (e.g. linting and testing with pre-commit hooks)
  • Empower it to check the work (e.g. "use npx <tool> to run a SQL query", "use use uvx pytest to run all the tests")
  • Don't PR code unless you're willing to maintain it
  • https://www.anthropic.com/engineering/claude-code-best-practices -- generally helpful
LLM Coding Z Talk, 2025-08-13

Demos

  • Fancy auto-complete/refactoring
  • Quick one-off code generation with new library feature
  • Real-world codebase understanding/refactoring
LLM Coding Z Talk, 2025-08-13

Q&A

Questions, comments, ideas?

LLM Coding Z Talk, 2025-08-13