Working with the agent
The agent is good at turning intent into working code. You'll get the best results by treating it like a capable teammate: be clear about what you want, then iterate.
Describe the outcome, not the implementation
Focus on what the app should do and how it should feel, and let the agent choose the how.
Instead of: "Create a React context, add a reducer, and wire up localStorage." Try: "Remember the user's theme choice between visits."
You can always get specific about the stack if you have a preference — but you don't have to.
Start broad, then refine
A good first prompt sketches the whole thing:
A personal finance dashboard: add transactions, categorize them, and show a monthly spending chart.
Then refine in follow-ups:
Make the chart a stacked bar by category.
Add a filter for the last 30 / 90 / 365 days.
Small, focused follow-ups are easier for the agent to get exactly right than one giant message with ten requirements.
Give it something to look at
If you have a design in mind, attach an image. The agent can see it and build to match. "Make it look like this" genuinely works.
When it self-corrects
You'll often see the agent run a build or start the app, hit an error, and immediately fix it. That's expected and healthy — it's checking its own work. Let it finish the loop; it usually resolves the issue without any input from you.
If it gets stuck or heads in the wrong direction, you have two good moves:
- Tell it. "That's not quite right — the total should update when I delete a row." Concrete feedback gets concrete fixes.
- Roll back. Restore an earlier checkpoint and try a different prompt.
Keep context tight
If you start a very different feature, it can help to say so explicitly ("Now let's switch to the settings page"). Clear scoping keeps the agent focused on the right part of your project.
Next: never lose work with Checkpoints and rollback.