ML Kit GenAI Prompt API

ML Kit GenAI Prompt API is Google's Android skill for sending natural-language requests to Gemini Nano on a device. It covers prompt design, structured output, prefix caching, system instructions, model availability, and lifecycle cleanup.

The install is small, but the runtime requirements are not. This is on-device AI with a device-gated model, a beta SDK, and APIs that are still changing. The skill is most useful when you treat it as a checklist for those constraints, not as a guarantee that a prompt will run on every Android phone.

Source and revision

What the skill actually does

The core file gives a short implementation path:

1. Prompt optimization. Use concise prompts, examples for in-context learning, XML/HTML delimiters for dynamic input, output constraints, and the Structured Output API when the response must be parsed.
2. Prefix caching. If the prompt is more than 200 words, separate a static prefix from the dynamic suffix and use the prefix caching API.
3. Lifecycle. Check FeatureStatus before the first inference. Download the model if it is DOWNLOADABLE. Call close() on the GenerativeModel when the activity, fragment, or view model is destroyed.
4. Structured output. Verify isStructuredOutputFeatureAvailable() before use, return a @Generable typed object instead of a string, and handle parse errors.

The six local references expand on prompt optimization, prompt design, prefix caching, getting started, structured output, and system instructions. The skill routes to them rather than duplicating everything.

Runtime constraints you must plan for

The version detail is a real drift signal: the SKILL.md says 1.0.0-beta4, while the get-started reference at the same commit shows 1.0.0-beta2. Treat the skill's prerequisite as the newer requirement and verify the current release in the Android documentation before pinning a version.

Illustrative implementation

This is a constructed example based on the skill's documented API, not a compiled or run app.

Task: parse a customer email into sender, title, and a short summary.

Before: the function returns a String, and the caller parses JSON by hand.

After:

@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 in 10 words or fewer")
    var summary: String = ""
)

suspend fun parseEmail(email: String): ParsedEmail? {
    val prompt = "Parse this email: $email"
    val baseRequest = GenerateContentRequest.Builder(TextPart(prompt)).build()
    val typedRequest = generateTypedContentRequest(baseRequest, ParsedEmail::class)
    val response = generativeModel.generateContent(typedRequest)
    return response.candidates.firstOrNull()?.response
}

The surrounding code still needs:

• A checkStatus() branch that handles UNAVAILABLE, DOWNLOADABLE, DOWNLOADING, and AVAILABLE.
• A fallback UI when the model is unavailable or the download fails.
• isStructuredOutputFeatureAvailable() before the typed request.
• ProGuard keep rules for the annotated class in release builds.
• close() in the owner's onDestroy or onCleared.

The skill's prompt-optimization reference also recommends delimiters such as <email>[content]</email> so the model can distinguish instructions from user data. That is a prompt-injection mitigation, not a guarantee.

When to use it

• You are adding on-device summarization, classification, entity extraction, or rewriting to an Android app.
• You need structured output instead of parsing free-form JSON.
• You have a long, repeated prompt prefix and want to test prefix caching.
• You need to check Gemini Nano availability and handle model download.
• You are reviewing an Android codebase that already uses ML Kit GenAI.

When not to use it

• You need a model that runs on every Android device. Gemini Nano availability is device-gated.
• You need a cloud model with larger context or stronger reasoning. This is on-device and constrained.
• You are building for iOS, web, or desktop.
• You need a stable, long-term API. The SDK and several features are beta or experimental.
• You need to process images with prefix caching. Prefix caching is text-only.
• You need a server-side key or a hosted endpoint. There is no API key in this workflow; the model is on the device.

Install

android skills add ml-kit-genai-prompt-api --project=.

Prerequisites and footprint

Prefix caching stores encrypted cache files and metadata, including the original prefix text, in the app's private storage. The reference says the cache count is managed by LRU and can be cleared with clearImplicitCaches().

Limitations

• Device availability is the hard gate. A device that does not support Gemini Nano cannot run the feature, regardless of minSdk.
• Beta and experimental APIs. Prefix caching is experimental, clearImplicitCaches() is experimental, and the Prompt API dependency is beta.
• Structured output is Kotlin-only. Java projects cannot use the typed output path described here.
• Version drift. The core file and the get-started reference disagree on the dependency version. Verify the current release.
• No credential story, but a download story. The absence of API keys does not mean zero network or storage cost.
• Prompt injection still applies. On-device content can contain instructions. Delimiters help; they are not a security boundary.
• No runtime test by us. We did not build the app or run it on a device.

Verify it yourself

1. Read the skill and all six references. Confirm the API level, dependency version, KSP version, and model requirements for your target devices.
2. Check the current Prompt API device-support list. Do not assume a flagship device from last year is supported.
3. Write a small screen that calls checkStatus() and renders every state, including UNAVAILABLE and DOWNLOADABLE. Test it on a device that lacks the model.
4. Test structured output with a deliberately invalid value to see how finishReason and GenAiException surface.
5. Add the ProGuard keep rule and test a release build, not only debug.
6. Verify close() is called and that a second screen does not reuse a closed model.

Review transparency

This is an original guide with short attributed code examples. It does not reproduce the full skill or its reference files.

Related

• Gemini Skills — Gemini CLI skills, not on-device Android
• Kotlin Rules — editor conventions
• Jetpack Compose Rules — UI conventions
• Prompt Optimization — model-agnostic prompt work
• How to Vet an Agent Skill — review before installing