Visual Intelligence

rshankras/visual-intelligence

Integrate your app with iOS Visual Intelligence for camera-based search and object recognition. Use when adding visual search capabilities.

5k tokens

context cost

the whole folder, loaded on every use

files

instructions only

copies elsewhere

how many repositories repackaged it

585

stars on the repo

on the repository, not the skill itself

Install

one command, takes just this skill from the repository

npx skills add https://github.com/rshankras/claude-code-apple-skills --skill visual-intelligence

The instruction itself

33 sections, as written by the author

Visual Intelligence

Integrate your app with iOS Visual Intelligence to let users find app content by pointing their camera at objects.

When This Skill Activates

User wants camera-based search in their app
User asks about visual search integration
User wants to surface app content in system searches
User needs to handle visual intelligence queries

Overview

Visual Intelligence lets users:

Point camera at objects or use screenshots
System identifies what they're looking at
Your app provides matching content
Results appear in system UI

Your app implements:

IntentValueQuery to receive search requests
AppEntity types for searchable content
Display representations for results

Platform Availability (WWDC26 297)

Visual Intelligence runs on iOS, iPadOS, and macOS — the same entities, query, and OpenIntent code works unchanged on all three. Handle both camera captures of physical objects (iOS) and screenshots of digital media (iPad/Mac) as input.
On Mac, the input pixel buffer can be much larger than what you'd encounter on iPhone — consider whether resizing is necessary before matching.

Quick Start

1. Import Frameworks

import VisualIntelligence
import AppIntents

2. Create App Entity

struct ProductEntity: AppEntity {
    var id: String
    var name: String
    var price: String
    var imageName: String

    static var typeDisplayRepresentation: TypeDisplayRepresentation {
        TypeDisplayRepresentation(
            name: LocalizedStringResource("Product"),
            numericFormat: "\(placeholder: .int) products"
        )
    }

    var displayRepresentation: DisplayRepresentation {
        DisplayRepresentation(
            title: "\(name)",
            subtitle: "\(price)",
            image: .init(named: imageName)
        )
    }

    // Deep link URL
    var appLinkURL: URL? {
        URL(string: "myapp://product/\(id)")
    }
}

3. Create Intent Value Query

struct ProductIntentValueQuery: IntentValueQuery {
    func values(for input: SemanticContentDescriptor) async throws -> [ProductEntity] {
        // Search using labels
        if !input.labels.isEmpty {
            return await searchProducts(matching: input.labels)
        }

        // Search using image
        if let pixelBuffer = input.pixelBuffer {
            return await searchProducts(from: pixelBuffer)
        }

        return []
    }

    private func searchProducts(matching labels: [String]) async -> [ProductEntity] {
        // Search your database using provided labels
        // Return matching products
    }

    private func searchProducts(from pixelBuffer: CVReadOnlyPixelBuffer) async -> [ProductEntity] {
        // Use image recognition on the pixel buffer
        // Return matching products
    }
}

SemanticContentDescriptor

The system provides this object with information about what the user is looking at.

Properties

| Property | Type | Description |

|----------|------|-------------|

| labels | [String] | Classification labels from Visual Intelligence |

| pixelBuffer | CVReadOnlyPixelBuffer? | Raw image data |

Usage Patterns

Label-based Search:

func values(for input: SemanticContentDescriptor) async throws -> [ProductEntity] {
    // Labels like "shoe", "sneaker", "Nike" etc.
    let labels = input.labels

    // Search your content using these labels
    return products.filter { product in
        labels.contains { label in
            product.tags.contains(label.lowercased())
        }
    }
}

Image-based Search:

func values(for input: SemanticContentDescriptor) async throws -> [ProductEntity] {
    guard let pixelBuffer = input.pixelBuffer else {
        return []
    }

    // Convert to CGImage for processing
    let ciImage = CIImage(cvPixelBuffer: pixelBuffer)
    let context = CIContext()

    guard let cgImage = context.createCGImage(ciImage, from: ciImage.extent) else {
        return []
    }

    // Use your ML model or image matching logic
    return await imageSearch.findMatches(for: cgImage)
}

Multiple Result Types

Use @UnionValue when your app has different content types.

Rules (WWDC26 297):

An app can have only ONE IntentValueQuery that accepts a SemanticContentDescriptor. All result types must flow through that single query — a @UnionValue enum with one case per entity type.
Every entity type in the union needs its own OpenIntent — without one, results of that type can't appear in image search.
Think beyond pixel matching: an album matched by image similarity can also surface the artist's nearby concerts — be creative about the type of content you return based on the context.

@UnionValue
enum SearchResult {
    case product(ProductEntity)
    case category(CategoryEntity)
    case store(StoreEntity)
}

struct VisualSearchQuery: IntentValueQuery {
    func values(for input: SemanticContentDescriptor) async throws -> [SearchResult] {
        var results: [SearchResult] = []

        // Search products
        let products = await productSearch(input.labels)
        results.append(contentsOf: products.map { .product($0) })

        // Search categories
        let categories = await categorySearch(input.labels)
        results.append(contentsOf: categories.map { .category($0) })

        return results
    }
}

Display Representations

Create compelling visual representations for search results.

Result Card Real Estate (WWDC26 297)

The search-result card gives about three lines of text for a title and subtitle, plus a thumbnail image — put the most important identifying info there (album name + artist).
With multiple results the sheet uses a two-column layout: if you initialize DisplayRepresentation with an image URL, serve a thumbnail-sized image, not the full-resolution asset — smaller images load faster.
Exception: a single result renders its image at the full width of the results sheet — don't over-shrink for that case.

// ❌ Full-res image URLs in DisplayRepresentation for multi-result responses
// ✅ Thumbnail-sized images (two-column sheet); full-width only when returning one result

Basic Display

var displayRepresentation: DisplayRepresentation {
    DisplayRepresentation(
        title: "\(name)",
        subtitle: "\(description)",
        image: .init(named: thumbnailName)
    )
}

With System Image

var displayRepresentation: DisplayRepresentation {
    DisplayRepresentation(
        title: "\(name)",
        subtitle: "\(category)",
        image: .init(systemName: "tag.fill")
    )
}

Rich Display

var displayRepresentation: DisplayRepresentation {
    DisplayRepresentation(
        title: LocalizedStringResource("\(name)"),
        subtitle: LocalizedStringResource("\(formatPrice(price))"),
        image: DisplayRepresentation.Image(named: imageName)
    )
}

On-Device Image Matching (WWDC26 297)

Whether you're searching on device or hitting a server, the same principles apply: return results fast and ranked.

Vision-framework pattern:

Pre-compute GenerateImageFeaturePrintRequest feature prints for your catalog; at query time, convert the pixel buffer via VideoToolbox (VTCreateCGImageFromCVPixelBuffer) and generate just one feature print to compare.
Filter with a maximum distance threshold to drop dissimilar results, sort ascending by distance so the best match is first, and cap the result count. Apple's sample signature: search(matching:limit: Int = 10, maxDistance: Double = 1.0).
Return [] when nothing matches or the pixel buffer is absent — the system handles displaying an empty response. ❌ Don't pad with weak matches.

// ❌ Compute feature prints at query time
// ✅ Pre-compute catalog prints; query = 1 print + threshold + sort + limit
let matches = catalogPrints
    .map { entry in (entry, entry.print.distance(to: queryPrint)) }
    .filter { $0.1 <= maxDistance }
    .sorted { $0.1 < $1.1 }
    .prefix(limit)

Vision offers more than feature prints for visual search: text extraction, barcode scanning, face detection, image classification (WWDC26 297).

Deep Linking

Enable users to open specific content from search results.

OpenIntent Rules (WWDC26 297)

Tapping a result runs your OpenIntent for that entity type, and its perform() runs as the app comes to the foreground:

Do navigation in perform(); defer heavy loading until after the view appears.
Take people straight to the content they selected — no intermediate screens.
✅ Reuse the OpenIntent from your existing App Intents adoption — you don't need a separate one just for Visual Intelligence. ❌ Duplicate per-feature OpenIntents.

// ❌ Heavy loading inside OpenIntent.perform (runs during foregrounding)
// ✅ Navigate only; load after the view appears; reuse one OpenIntent everywhere

URL-based Deep Links

struct ProductEntity: AppEntity {
    // ... other properties

    var appLinkURL: URL? {
        URL(string: "myapp://product/\(id)")
    }
}

Handle in App

@main
struct MyApp: App {
    var body: some Scene {
        WindowGroup {
            ContentView()
                .onOpenURL { url in
                    handleDeepLink(url)
                }
        }
    }

    func handleDeepLink(_ url: URL) {
        guard url.scheme == "myapp" else { return }

        switch url.host {
        case "product":
            let id = url.lastPathComponent
            navigationState.showProduct(id: id)
        default:
            break
        }
    }
}

"More Results" Button

Provide access to additional results beyond the initial set.

struct ViewMoreProductsIntent: AppIntent, VisualIntelligenceSearchIntent {
    static var title: LocalizedStringResource = "View More Products"

    @Parameter(title: "Semantic Content")
    var semanticContent: SemanticContentDescriptor

    func perform() async throws -> some IntentResult {
        // Store search context for your app
        SearchContext.shared.currentSearch = semanticContent.labels

        // Return empty result - system will open your app
        return .result()
    }
}

semanticContentSearch Schema (WWDC26 297)

The schema-based form: conform to .visualIntelligence.semanticContentSearch and the system supplies the semanticContent property automatically:

@AppIntent(schema: .visualIntelligence.semanticContentSearch)
struct SemanticContentSearchIntent: AppIntent {
    static let openAppWhenRun: Bool = true

    var semanticContent: SemanticContentDescriptor

    func perform() async throws -> some IntentResult {
        let results = try await library.search(matching: semanticContent)
        await MainActor.run { AppState.shared.openSearch(with: results) }
        return .result()
    }
}

Rules (WWDC26 297): pre-populate the in-app search view from the captured context (never a blank screen), and use it to expose what the Visual Intelligence sheet can't — filters, categories, the full depth of your content.

Complete Example

import SwiftUI
import AppIntents
import VisualIntelligence

// MARK: - Entities

struct RecipeEntity: AppEntity {
    var id: String
    var name: String
    var cuisine: String
    var prepTime: String
    var imageName: String

    static var typeDisplayRepresentation: TypeDisplayRepresentation {
        TypeDisplayRepresentation(
            name: LocalizedStringResource("Recipe"),
            numericFormat: "\(placeholder: .int) recipes"
        )
    }

    var displayRepresentation: DisplayRepresentation {
        DisplayRepresentation(
            title: "\(name)",
            subtitle: "\(cuisine) · \(prepTime)",
            image: .init(named: imageName)
        )
    }

    var appLinkURL: URL? {
        URL(string: "recipes://recipe/\(id)")
    }
}

// MARK: - Intent Value Query

struct RecipeVisualSearchQuery: IntentValueQuery {
    @Dependency var recipeStore: RecipeStore

    func values(for input: SemanticContentDescriptor) async throws -> [RecipeEntity] {
        // Use labels to find recipes
        // Labels might include: "pasta", "tomato", "Italian", etc.
        let matchingRecipes = await recipeStore.search(
            ingredients: input.labels,
            limit: 15
        )

        return matchingRecipes.map { recipe in
            RecipeEntity(
                id: recipe.id,
                name: recipe.name,
                cuisine: recipe.cuisine,
                prepTime: recipe.prepTimeFormatted,
                imageName: recipe.thumbnailName
            )
        }
    }
}

// MARK: - More Results Intent

struct ViewMoreRecipesIntent: AppIntent, VisualIntelligenceSearchIntent {
    static var title: LocalizedStringResource = "View More Recipes"

    @Parameter(title: "Semantic Content")
    var semanticContent: SemanticContentDescriptor

    func perform() async throws -> some IntentResult {
        // Save search context
        await MainActor.run {
            RecipeSearchState.shared.searchTerms = semanticContent.labels
        }
        return .result()
    }
}

// MARK: - Recipe Store

@Observable
class RecipeStore {
    private var recipes: [Recipe] = []

    func search(ingredients: [String], limit: Int) async -> [Recipe] {
        recipes
            .filter { recipe in
                ingredients.contains { ingredient in
                    recipe.ingredients.contains { recipeIngredient in
                        recipeIngredient.lowercased().contains(ingredient.lowercased())
                    }
                }
            }
            .prefix(limit)
            .map { $0 }
    }
}

Best Practices

Performance

Return results quickly (< 1 second)
Limit initial results (10-20 items)
Use "More Results" for additional content
Cache search indexes for fast lookup

func values(for input: SemanticContentDescriptor) async throws -> [ProductEntity] {
    // Limit results for quick response
    let results = await search(input.labels)
    return Array(results.prefix(15))
}

Relevance

Prioritize exact matches
Consider context (location, time)
Filter low-confidence matches

func values(for input: SemanticContentDescriptor) async throws -> [ProductEntity] {
    let results = await search(input.labels)

    // Sort by relevance score
    return results
        .filter { $0.relevanceScore > 0.5 }
        .sorted { $0.relevanceScore > $1.relevanceScore }
        .prefix(15)
        .map { $0 }
}

Quality Representations

Use clear, concise titles
Include helpful subtitles
Provide relevant thumbnails
Localize all text

var displayRepresentation: DisplayRepresentation {
    DisplayRepresentation(
        title: LocalizedStringResource(stringLiteral: name),
        subtitle: LocalizedStringResource(
            stringLiteral: "\(category) · \(formattedPrice)"
        ),
        image: .init(named: thumbnailName)
    )
}

Receiving Visual Intelligence Data (WWDC26 297)

Two integration directions: your app provides results (everything above), and your app receives data Visual Intelligence writes into shared system stores:

| Visual Intelligence system action | Store | Your app reads via |

|-----------------------------------|-------|--------------------|

| Create calendar events — including multiple events at once | EventKit | EKEventStore |

| Add to contacts | Contacts | CNContactStore |

| Log medical device readings (blood pressure monitors, glucose meters, weight scales) | HealthKit | HKHealthStore |

If your app already reads from these stores, Visual Intelligence becomes a source of input automatically — zero VI-specific code. One requirement: observe change notifications so VI-created data appears without a relaunch. EventKit pattern from Apple's sample: requestFullAccessToEvents() → fetch with a predicate (a 90-day window) → observe .EKEventStoreChanged notifications and refetch.

Testing

Build and run on physical device
Open Camera or take screenshot
Activate Visual Intelligence
Point at objects relevant to your app
Verify results appear
Test tapping results opens your app correctly
On iPad and Mac, test with screenshots of digital media — the primary entry point there (WWDC26 297)

Checklist

[ ] Import VisualIntelligence and AppIntents
[ ] Create AppEntity types for searchable content
[ ] Implement IntentValueQuery
[ ] Handle both labels and pixelBuffer
[ ] Create DisplayRepresentation for each entity
[ ] Implement deep linking URLs
[ ] Handle URLs in app with onOpenURL
[ ] Add "More Results" intent if needed — pre-populated, never a blank search screen
[ ] Only one SemanticContentDescriptor-accepting IntentValueQuery in the app (WWDC26 297)
[ ] OpenIntent per entity type; navigation only in perform(), heavy loading deferred (WWDC26 297)
[ ] Thumbnail-sized display images for multi-result responses (WWDC26 297)
[ ] Return [] when nothing matches — no weak-match padding (WWDC26 297)
[ ] Feature prints pre-computed, results filtered by max distance and sorted best-first (WWDC26 297)
[ ] Test on physical device
[ ] Optimize for performance (< 1s response)
[ ] Localize display text

References

How to use it

Copy the folder

Take rshankras/visual-intelligence from the repository into ~/.claude/skills for personal use, or into .claude/skills inside a project.

Check the name does not clash

The agent identifies a skill by the name field in its header. Two skills with the same name cannot sit side by side — one of them will be ignored.