Skip to content
Verified

Prompting patterns

What to say to the AI App Builder, and how scope, mode and context change the answer.

The AI App Builder reads more from the project than from your words alone. These patterns come from what the panel actually does with a message, so the same sentence can land differently depending on three things you set before you type it.

Scope decides what it is allowed to touch

General scope means the whole open project is fair game. Node scope means only whatever is selected in the tree, useful the moment you want a change contained to one component instead of risking a project-wide read. Select the target first, then describe the change, rather than naming the target inside the sentence.

Switching scope starts a new chat

General and Node are not the same conversation. Pick the scope before you start describing the change, not partway through.

Mode decides whether it asks first

Auto applies a change end to end. Plan researches and proposes an approach without touching anything, then waits. Reach for Plan when a request could land several different ways (which table gets the new field, whether a screen becomes a new component or an addition to an existing one), so you approve the shape before it is built rather than undoing it after.

Context chips already say what it can see

The composer shows what the next message will use: the open program, the tree selection when it differs from what is open, and the selected UI element when the UI tab is open. Do not repeat what a chip already states, an instruction is redundant the moment its subject is already attached. Toggle a chip off for a single message when you want the assistant to ignore it, for example a UI element still selected for an unrelated reason.

Build an agent by describing the job, not a prompt

The ai_agent builder does not take a system prompt, it takes a goal: What should this agent do?, answered in a couple of sentences. From there Studio proposes the tools, the instructions and any other agents the job needs, and you see the full plan before anything is built. A goal that leaves real gaps (who uses it, what it should draw on, what it should produce) gets a short list of clarifying questions first, each with suggested answers to pick from or edit. Attach reference material the same way you would give a person more than a title: the document it must produce, a screenshot of the report, the price list it has to read.

When you are prompting an agent, not a person

Point an MCP-connected coding agent at a project instead of typing into the panel yourself, and the prompt becomes instructions to that agent rather than a message to Studio's own assistant. Say what you are building in terms the Studio LLM guide already gives it: name the prog type (component, table, get_data, javascript...), and let the agent's own required workflow do the rest, read the matching guide topic, fetch a working example, create, validate, and only then save. An agent that already has the guide needs the goal, not a tutorial on xu-bind syntax.