Working with the AI

The AI in Vybe is more than a code generator. It is a collaborative partner that can build, debug, explain, and iterate on your app. Getting the most out of it means understanding how it works and adapting your workflow accordingly.

---

​

Agent mode vs. Ask mode

Vybe has two interaction modes, and switching between them at the right time makes a real difference.

​

Agent mode (default)

In Agent mode, the AI can create, edit, and delete files, run database queries, install packages, and connect integrations. Use it whenever you want the AI to make changes to your app.

Add a new page that lists all orders from the last 30 days with status filters.

​

Ask mode

In Ask mode, the AI can read your code and answer questions, but it cannot make any changes. Use it when you want to understand your app without risking unintended modifications.

How does the authentication flow work in this app?

What database tables are currently defined?

Explain what the /api/customers route does.

---

​

Choosing the right model

Different models have different strengths. Picking the right one for the task at hand saves time and credits. A practical workflow:

1. Start with Opus 4.5 when building new features or tackling complex logic
2. Switch to Sonnet 4.5 for UI adjustments and straightforward changes
3. Use Haiku 4.5 for quick questions and minor text edits
4. Escalate to Opus 4.5 - High when the standard model struggles with a problem
5. Try GPT-5.1 or Gemini 3 Pro if you want a different approach to a tricky problem

For the full model reference, see AI Models Reference.

---

​

Understanding locked files

Certain system files in your app are locked and cannot be edited by you or the AI:

• package.json — manages dependencies
• next.config.ts — Next.js configuration
• tsconfig.json — TypeScript configuration
• Middleware configuration files

These files are locked to keep your app’s infrastructure stable and compatible with the Vybe platform. The AI is aware of these restrictions and will not attempt to modify them.

---

​

When the AI asks you questions

Sometimes the AI needs more information before it can proceed. When this happens, it presents interactive forms with options to choose from or fields to fill in. For example, if you ask “add user authentication,” the AI might ask:

• What fields should the user profile include?
• Should users sign up with email/password or OAuth?
• What pages should be protected?

These questions help the AI make better decisions. Take the time to answer thoughtfully — it results in code that matches your requirements more closely. You can always provide custom input beyond the suggested options if none of the choices fit your needs.

---

​

Handling errors

Errors are a normal part of building apps. Here is how to handle them efficiently:

​

Preview errors

If the live preview shows an error, click the error message to send it directly to the AI. The AI can read the error, trace it to the source, and fix it — often in a single step.

​

Build and type errors

Before deploying, ask the AI to run type checks:

Run a type check on the project and fix any errors.

This catches issues like missing imports, incorrect types, and broken API route signatures before they reach production.

​

When the AI gets stuck

If the AI produces incorrect output or seems confused:

1. Rephrase your request — try describing the same goal from a different angle
2. Reference specific files — “Update src/app/customers/page.tsx to add a delete button for each row”
3. Break it down — split a complex request into two or three simpler steps
4. Switch models — Opus 4.5 - High or a different provider may handle the problem better
5. Start a new thread — if the conversation has become cluttered, a fresh thread can help

---

​

Thread management

Each app can have multiple conversation threads. Use threads to keep your work organized:

• One feature per thread works well — it keeps the conversation focused and the context relevant
• Create a new thread when switching to a different topic or feature area
• Return to old threads to continue work — the AI retains the full conversation history
• Use descriptive names for threads so you can find them later

---

​

Getting the most out of each interaction

​

Give context upfront

Instead of making the AI discover context through trial and error, provide it in your prompt:

The customers page currently shows a table with name and email. Add a “Plan” column that reads from the plan_type field in the customers table.

​

Be explicit about behavior

Describe not just what something should look like, but how it should behave:

When the user clicks Delete, show a confirmation dialog. If they confirm, call the delete API and show a success toast. If the API fails, show an error toast and keep the row in the table.

​

Use follow-up prompts for refinement

The AI responds better to iterative refinement than to monolithic prompts:

1. “Build a customer list page with a data table.”
2. “Add search and filter by plan type.”
3. “Add a detail view that opens when you click a row.”
4. “Add edit and delete functionality to the detail view.”

Each step builds on the previous one, and you can course-correct along the way.

---

​

What’s next

• Learn prompting techniques: Prompting Best Practices
• Understand platform constraints: Platform Limitations
• Troubleshoot common issues: Troubleshooting