Prompting best practices

Updated: Feb 27, 2026

This guide helps you write effective prompts when using AI tools connected through the MCP for Meta Horizon OS developers. Better prompts lead to faster, more accurate results, whether you’re querying documentation, debugging a device, analyzing performance, or searching for 3D assets.

Core principles

State your platform, SDK, and context upfront

The MCP server supports multiple SDKs and development targets. Specifying yours eliminates ambiguity and gets you relevant answers on the first try.

Instead of: “How do I render passthrough?”
Write: “How do I enable passthrough rendering in Unity using the Meta XR SDK on Quest 3?”

Instead of: “Set up hand tracking”
Write: “Walk me through implementing hand tracking with the Meta Spatial SDK for an Android-native Quest app.”

Instead of: “Fix my shader”
Write: “My custom URP shader isn’t rendering correctly on Quest 2. Here’s the shader code: ...”

Include the right amount of context

Give the LLM enough information to act without overloading it. Include what’s relevant, leave out what isn’t.

For debugging, include:

The error message or logcat output

What you were doing when it occurred

Your target device and OS version

The SDK and engine version you’re using

For documentation discovery, include:

The specific feature or API you’re asking about

Your development framework (Unity, Unreal, Native)

For performance analysis, include:

The session ID or trace file name

The time range or frame range you care about

What metric or behavior you’re investigating

Use direct, imperative language

Tell the LLM what to do. Avoid hedging or open-ended phrasing when you know what you need.

Instead of: “Can you maybe check the logs?”
Write: “Pull the last 500 lines of logcat filtered to tag Unity at warning level and above.”

Instead of: “I was wondering about the docs”
Write: “Search the documentation for scene anchors in Unity.”

Instead of: “Something seems slow”
Write: “Load trace session_42, query thread states between timestamps 1000000 and 5000000, and identify which threads spent the most time in runnable state.”