This lesson's interactive features are locked, please to keep using them
Workflow
The tasks you work on will vary from single-line changes to multi-file features that require new infrastructure to be deployed.
That means sometimes your workflow will be as simple as one-shotting a single-sentence prompt. Other times, it will be a full 30-message back and forth with the agent before you even let it change a single line. And sometimes, you'll still execute the change (gasp) by hand.
That said, in most cases, the following basic workflow will serve you well. Have the agent:
- Explore the codebase and constraints
- Plan the change if it's non-trivial
- Implement the smallest correct version
- Verify with real feedback loops
Why This Works
Planning and implementing are obvious steps – and hopefully, verifying is too. The most overlooked part in my experience is the first step: exploring the codebase, the constraints, documentation, and more.
Without really good context, the agent's plan is going to be suboptimal. The "way to build a notes feature" is going to look very different for a local-first app like Obsidian than it will for a realtime web-based app like Google Docs. Don't be afraid to force the agent to go read!
And exploring shouldn't be limited to the codebase. Agents have the webfetch tool! Make it do research about common approaches, which libraries exist, and anything else you think might be relevant.