SKILL.md

name	mobile-ai-features
description	Integrate AI APIs (OpenAI, Anthropic, Google AI) into a React Native/Expo app. Covers vision, text generation, and audio transcription with secure API key management. Use when the user wants to add AI-powered features to their mobile app.
standards-version	1.9.0

Mobile AI Features

Trigger

Use this skill when the user:

• Wants to add AI-powered features to a mobile app
• Needs to send images to a vision API (GPT-4o, Claude, Gemini)
• Wants text generation or chatbot functionality
• Needs audio transcription (Whisper)
• Asks about securing API keys in a mobile app
• Mentions "AI", "GPT", "Claude", "Gemini", "vision", "transcription", "chatbot", or "LLM"

Required Inputs

• AI provider: OpenAI, Anthropic, or Google AI
• Features needed: vision, text generation, audio transcription, or a combination
• Backend strategy: edge function (Cloudflare Workers, Vercel Edge), server (Express, Fastify), or serverless (AWS Lambda, Supabase Edge Functions)

Workflow

1. Never bundle API keys in the app. Mobile app binaries can be decompiled. Anyone with the APK or IPA can extract hardcoded keys.

   Instead, use a backend proxy:

   Mobile App  -->  Your Backend  -->  AI Provider
                    (holds API key)
   

   The backend holds the API key. The mobile app authenticates with your backend using user auth tokens.

2. Set up a minimal backend proxy. Example with Cloudflare Workers:

   export default {
     async fetch(request: Request): Promise<Response> {
       const { prompt, image } = await request.json();
   
       const response = await fetch("https://api.openai.com/v1/chat/completions", {
         method: "POST",
         headers: {
           "Content-Type": "application/json",
           Authorization: `Bearer ${env.OPENAI_API_KEY}`,
         },
         body: JSON.stringify({
           model: "gpt-4o",
           messages: [
             {
               role: "user",
               content: image
                 ? [
                     { type: "text", text: prompt },
                     { type: "image_url", image_url: { url: image } },
                   ]
                 : prompt,
             },
           ],
           max_tokens: 1024,
         }),
       });
   
       return response;
     },
   };

   Deploy this and point your app at its URL.

3. Create an API client in the app. In lib/ai.ts:

   const API_BASE = "https://your-worker.your-domain.workers.dev";
   
   interface AIResponse {
     text: string;
     usage: { prompt_tokens: number; completion_tokens: number };
   }
   
   export async function generateText(
     prompt: string,
     authToken: string,
   ): Promise<AIResponse> {
     const res = await fetch(`${API_BASE}/generate`, {
       method: "POST",
       headers: {
         "Content-Type": "application/json",
         Authorization: `Bearer ${authToken}`,
       },
       body: JSON.stringify({ prompt }),
     });
   
     if (!res.ok) {
       const error = await res.text();
       throw new Error(`AI request failed: ${res.status} ${error}`);
     }
   
     return res.json();
   }
   
   export async function analyzeImage(
     imageBase64: string,
     prompt: string,
     authToken: string,
   ): Promise<AIResponse> {
     const res = await fetch(`${API_BASE}/vision`, {
       method: "POST",
       headers: {
         "Content-Type": "application/json",
         Authorization: `Bearer ${authToken}`,
       },
       body: JSON.stringify({
         prompt,
         image: `data:image/jpeg;base64,${imageBase64}`,
       }),
     });
   
     if (!res.ok) throw new Error(`Vision request failed: ${res.status}`);
     return res.json();
   }

4. Use vision with expo-camera. Capture a photo and send it to the vision API:

   import { CameraView } from "expo-camera";
   import { useRef } from "react";
   import { analyzeImage } from "@/lib/ai";
   
   const cameraRef = useRef<CameraView>(null);
   
   async function captureAndAnalyze() {
     if (!cameraRef.current) return;
   
     const photo = await cameraRef.current.takePictureAsync({
       base64: true,
       quality: 0.5,
     });
   
     if (!photo?.base64) return;
   
     const result = await analyzeImage(
       photo.base64,
       "Describe what you see in this image.",
       userAuthToken,
     );
   
     console.log(result.text);
   }

   Use quality: 0.5 or lower to reduce payload size. A full-resolution photo can be 5MB+ in base64.

5. Streaming responses. For chat UIs, stream tokens as they arrive:

   export async function streamText(
     prompt: string,
     authToken: string,
     onToken: (token: string) => void,
   ): Promise<void> {
     const res = await fetch(`${API_BASE}/stream`, {
       method: "POST",
       headers: {
         "Content-Type": "application/json",
         Authorization: `Bearer ${authToken}`,
       },
       body: JSON.stringify({ prompt, stream: true }),
     });
   
     if (!res.ok) throw new Error(`Stream failed: ${res.status}`);
     if (!res.body) throw new Error("No response body");
   
     const reader = res.body.getReader();
     const decoder = new TextDecoder();
   
     while (true) {
       const { done, value } = await reader.read();
       if (done) break;
       const chunk = decoder.decode(value, { stream: true });
       onToken(chunk);
     }
   }

   Usage in a component:

   const [response, setResponse] = useState("");
   
   async function handleSend() {
     setResponse("");
     await streamText(prompt, authToken, (token) => {
       setResponse((prev) => prev + token);
     });
   }

6. Audio transcription. Record audio with expo-av and send to Whisper:

   npx expo install expo-av

   import { Audio } from "expo-av";
   
   const [recording, setRecording] = useState<Audio.Recording | null>(null);
   
   async function startRecording() {
     await Audio.requestPermissionsAsync();
     await Audio.setAudioModeAsync({ allowsRecordingIOS: true });
     const { recording } = await Audio.Recording.createAsync(
       Audio.RecordingOptionsPresets.HIGH_QUALITY,
     );
     setRecording(recording);
   }
   
   async function stopAndTranscribe() {
     if (!recording) return;
     await recording.stopAndUnloadAsync();
     const uri = recording.getURI();
     if (!uri) return;
   
     const formData = new FormData();
     formData.append("file", {
       uri,
       type: "audio/m4a",
       name: "recording.m4a",
     } as any);
   
     const res = await fetch(`${API_BASE}/transcribe`, {
       method: "POST",
       headers: { Authorization: `Bearer ${authToken}` },
       body: formData,
     });
   
     const { text } = await res.json();
     console.log("Transcription:", text);
   }

Key References

• OpenAI API: Vision
• Anthropic API: Vision
• Google AI: Gemini
• expo-av: Audio recording
• OpenAI Whisper API

Example Interaction

User: "I want to take a photo and have AI describe what's in it."

Agent:

1. Sets up expo-camera with mobile_installDependency
2. Adds camera permission with mobile_addPermission
3. Scaffolds an AI client with mobile_integrateAI (provider: openai, features: vision)
4. Creates a camera screen that captures a photo and sends base64 to the vision endpoint
5. Displays the AI description below the photo preview
6. Warns about not bundling API keys and recommends a backend proxy

MCP Usage

Step	MCP Tool	Description
Scaffold AI client	mobile_integrateAI	Create API client file with provider config and TypeScript types
Install expo-camera	mobile_installDependency	For vision features that use the camera
Install expo-av	mobile_installDependency	For audio recording and transcription
Add permissions	mobile_addPermission	Add camera/microphone permission rationale
Check build	mobile_checkBuildHealth	Verify project builds after adding native modules

Common Pitfalls

1. Bundling API keys in the app - This is the most common and most dangerous mistake. Mobile apps can be reverse-engineered. Always proxy through a backend you control.
2. Sending full-resolution images - A 12MP photo is 5-10MB in base64. Use quality: 0.3-0.5 in takePictureAsync and resize before sending. This also reduces API costs.
3. No error handling for network failures - Mobile apps lose connectivity. Wrap AI calls in try/catch, show user-friendly errors, and implement retry with exponential backoff.
4. Ignoring costs - Vision API calls cost more than text. Show users an estimate or add rate limiting on your backend. GPT-4o vision is roughly $0.01-0.03 per image.
5. Blocking the UI during AI calls - AI responses can take 2-10 seconds. Show a loading indicator and use streaming for chat interfaces.
6. Not setting a timeout - Add AbortController with a 30-second timeout to prevent indefinite hangs on slow connections.

See Also

• Mobile Camera Integration - capture photos to send to vision APIs
• Mobile Permissions - handle camera and microphone permissions