ml-kit-genai-prompt-api — Claude Code skill for on-device Prompt API

---
name: ml-kit-genai-prompt-api
description: Analyzes Android codebases to implement ML Kit GenAI Prompt API. Use
  this skill to send natural language requests on-device to Gemini Nano, use structured
  output with Prompt API, implement prefix caching, optimize the current prompt, or
  apply best practices."
license: Complete terms in LICENSE.txt
metadata:
  author: Google LLC
  last-updated: '2026-09-03'
  keywords:
  - ML Kit
  - Prompt API
  - Structured Output
  - Prefix Caching
  - Gemini Nano
---

This skill provides step-by-step guidance for integrating and optimizing the ML
Kit GenAI Prompt API in Android apps.

## Prerequisites

- Android API level must be 26 or higher. If `minSdk` is below 26, update it to 26.
- Add the ML Kit GenAI Prompt API dependency (`com.google.mlkit:genai-prompt`) to the app-level `build.gradle` file, with version at least `1.0.0-beta4`.
- If `com.google.mlkit:genai-schema-compiler` dependency is used and KSP plugin version is below 2.3.6, update it to 2.3.6.

## Detailed steps

### 1. Prompt optimization

To optimize prompts for use with the ML Kit Prompt API, follow the
[prompt optimization guide](https://developer.android.com/agents/skills/device-ai/prompt-api/references/prompt-optimization).

### 2. Prefix caching optimization

If the prompt is more than 200 words, implement the [prefix caching API](https://developer.android.com/agents/skills/device-ai/prompt-api/references/prefix-caching).

### 3. Lifecycle and best practices

- The model must be fully downloaded and available before calling the first inference. Follow the guide on [implementing a generative model](https://developer.android.com/agents/skills/device-ai/prompt-api/references/get-started) to check that the `FeatureStatus` of a model is `AVAILABLE` before making an inference.
- Release ML Kit instances by calling `close()` when an `Activity`,
  `Fragment`, or `ViewModel` is destroyed. Example:


  ```kotlin
  // Instantiating model in activity, fragment, or ViewModel
      val generativeModel = Generation.getClient()

  // When activity, fragment, or ViewModel is destroyed
      generativeModel.close()
  ```

  <br />

### 4. Structured output

When implementing or refactoring a prompt to use structured output, follow
these rules:

1. **Check for API availability:** Verify Structured Output feature is available on the device with `isStructuredOutputFeatureAvailable()` before using it. Refer to the [Structured Output API guide](https://developer.android.com/agents/skills/device-ai/prompt-api/references/structured-output) for full instructions.
2. **Return type:** Return the `@Generable` typed object from the function
   signature instead of a `String` or JSON string.

   For example:

       fun parseEmail(email: String): String {
           ... 
       }

   should be refactored to:

       fun parseEmail(email: String): ParsedEmail? {
           ...
       }

3. **Example:**

   This is the example code before refactoring:


   ```kotlin
   suspend fun parseEmail(email: String): String {
       val parseEmailPrompt = "Parse this email and return the sender, title, and short summary of the email less than 10 words: "

       val parsedEmail = generativeModel.generateContent(parseEmailPrompt + email)

       return parsedEmail.candidates[0].text
   }
   ```

   <br />

   This is the example code after using Structured Output API:


   ```kotlin
   @Generable
   data class ParsedEmail(
       @Guide(description = "Sender of the email")
       var sender: String = "",

       @Guide(description = "Title of the email")
       var title: String = "",

       @Guide(description = "Summary of the email less than 10 words")
       var summary: String = ""
   )

   suspend fun parseEmail(email: String): ParsedEmail? {
       val parseEmailPrompt =
           "Parse this email: $email"

       val baseRequest = GenerateContentRequest.Builder(TextPart(parseEmailPrompt)).build()
       val typedRequest = generateTypedContentRequest(baseRequest, ParsedEmail::class)
       val typedResponse = generativeModel.generateContent(typedRequest)
       return typedResponse.candidates[0].response
   }
   ```

   <br />