How to use Startbase

Best practices

Leverage knowledge files, plan mode, version control, and structured prompting to ship faster and sidestep common mistakes.

Whether you just signed up or have built a dozen projects, these habits will help you get more out of every session and avoid the pitfalls that slow builders down.

Start with a knowledge file

Your project's knowledge file is attached to every prompt the AI receives. Think of it as a living product brief: it keeps the agent aligned with your vision even across long conversations.

What to include
Product vision and goals, target audience and personas, key features and expected behaviors, design guidelines (colors, typography, spacing), and role-specific rules (e.g. Admin vs. User).

You can auto-generate a knowledge file by asking: "Generate a knowledge file for my project based on the features I've already built."

Write better prompts

Treat the AI like a skilled collaborator: the more context you share, the closer the first result will be to what you actually want.

Reference exact pages or routes (e.g. /dashboard, /settings/profile).
Attach screenshots when describing bugs or UI issues.
Tell the AI what NOT to change: "Do not edit Layout.tsx or the shared nav."
Repeat critical instructions across prompts; context can fade in long threads.
Avoid bundling five changes into one message. One prompt, one feature.

Vague prompts produce vague results. "Make it look better" gives the AI almost nothing to work with. "Use a dark theme with 8px rounded cards and a gradient hero from indigo-900 to black" delivers on the first try.

❌  "Build me an app"
✅  "Build a recipe-sharing app with user profiles,
    a search bar, and a favorites list. Use Tailwind
    with a warm color palette and rounded cards."

Use Plan mode early and often

Plan mode lets the AI reason about your project, exploring files, database schemas, and dependencies, without writing any code until you say "go."

When to switch
After two failed "Try to Fix" attempts, when debugging database or auth issues, when planning a new feature, or whenever you want the agent to think before acting.

Many experienced builders spend 60-70% of their time in Plan mode and only click "Implement the plan" once they are fully satisfied with the approach.

Use checkpoints and GitHub wisely

Every agent run creates a checkpoint. Use Project history to bookmark stable states mentally and roll back confidently.

After every working feature, note where you are; the checkpoint is already saved for you.
After a bug: compare behavior against the previous checkpoint. "What changed since the last run?"
If too much has broken, restore the last good checkpoint and try again with better prompts.
Push to GitHub for an external copy of any state you want to keep forever.

Start fresh when all else fails

Sometimes rebuilding with better prompts takes a fraction of the time the first attempt took. Restore an early checkpoint, or create a new project and bring your best prompts with you. Many builders find the second pass dramatically faster.

Useful when you are stuck in a loop of cascading errors.
Keep the original project as a reference while you rebuild.
Fork any community project (including your own published one) to get a clean copy to experiment on.