Demos
Demo prompts
24 prompts, word for word. Each one was run through SiliconDevKit to make a real Mac app. Read the prompt, then open the app page to download it.
Starter demos
The five prompts free accounts can build.
Prompt
Build a macOS app called Agent Swarm. It detects which AI command-line tools are installed on this Mac: claude, gemini and hermes. Check common install folders (/opt/homebrew/bin, /usr/local/bin, ~/.local/bin, ~/.npm-global/bin, ~/.hermes/bin) and also ask a login shell (zsh -lc 'command -v <tool>'), because apps opened from Finder do not inherit the shell PATH. Show each tool with a green or grey status dot. The user types one task; the app sends it to every detected tool at the same time, non-interactively with these commands: claude -p "<task>", gemini -p "<task>", hermes chat -q "<task>". Run them with Process using the login shell so each tool's own login is used, and stream each tool's output into its own side-by-side column. Show a clear 'not installed' message for missing tools, a Stop button, and a Copy button per column. Do not ask for API keys.
Prompt
Build a menu bar app called ClipShelf that remembers the last 50 things I copied. Show them in a popover list, newest first, with a search field. Clicking an item copies it back to the clipboard. Poll NSPasteboard for changes, skip duplicates, keep it text-only, and add a Clear all button and a Quit button. No Dock icon.
Prompt
Build a menu bar app called AppPulse that shows total CPU use in the menu bar. Its popover shows a live list of running apps with their CPU percent and memory, grouped by app instead of by process, sorted by CPU, refreshed every 2 seconds. Use NSWorkspace for app names and icons and the ps command for the numbers. Add a Quit button. No Dock icon.
Prompt
Build a Markdown notes app called Scribe. A sidebar lists notes saved as .md files in ~/Documents/Scribe (create the folder on first launch). The main area has a text editor with a live rendered Markdown preview side by side, autosave, and buttons to create, rename and delete notes. Add an Always on top toggle in the toolbar.
Prompt
Build an app called Dither Dunes: a window that renders a slowly animating black-and-white dithered landscape of rolling dunes with a sun, using SwiftUI Canvas and ordered Bayer dithering at a chunky pixel scale, at 30 fps. Add toolbar buttons for pause and full screen, and a palette picker with mono, amber and green.
Vision
Prompt
Build a menu bar app called Cutout that removes image backgrounds on-device. Clicking the menu bar icon opens a popover with a large drop zone that accepts image files dragged from Finder or the browser, plus a Choose File button and paste from the clipboard. Use the Vision framework: VNGenerateForegroundInstanceMaskRequest on macOS 14 and later (isolate all subjects or tap an object to keep only that one), and VNGeneratePersonSegmentationRequest as the fallback on macOS 13. Show the result as a cutout on a checkerboard background, with a before and after toggle, an edge feather slider and a background color picker (transparent, white, black, custom). Add buttons to copy the cutout PNG to the clipboard and to save it as a transparent PNG. Everything runs locally, nothing is uploaded.
Prompt
Build a menu bar app called TextSnap that turns screenshots into clean text. Clicking the menu bar icon opens a floating panel. Pressing Cmd+V in it (or dropping an image, or clicking Paste from clipboard) reads an image from the pasteboard and runs the Vision framework text recognizer (VNRecognizeTextRequest, accurate level, language correction on). Use the bounding boxes to rebuild the layout: detect rows and columns and offer an output mode picker with Markdown (headings, lists and paragraphs inferred from text size and position), Markdown table, TSV for spreadsheets, Plain text, and LaTeX (a best-effort conversion of math symbols, fractions written as a/b, superscripts and Greek letters). Show a Raw tab with monospaced text and a Preview tab with the rendered result, and a Copy button that puts the result on the clipboard. Everything runs locally.
Prompt
Build a macOS app called Bokeh Studio. Drop a photo onto the window (or use Open). Build a depth estimate on-device: use the embedded depth or disparity map of a Portrait mode photo when the file has one (read it with ImageIO auxiliary data), otherwise use a Vision subject mask (VNGenerateForegroundInstanceMaskRequest on macOS 14+, VNGeneratePersonSegmentationRequest on macOS 13) blended with a soft vertical gradient. Add two modes. Bokeh: an aperture slider (f/1.4 to f/16) and a focus point you set by clicking the photo, rendering a depth-weighted lens blur with Core Image (CIMaskedVariableBlur or CIDepthBlurEffect). Parallax: the foreground and background layers shift in opposite directions as the mouse moves over the image, for a live 3D look, with a strength slider. Add a Save as JPEG button for the bokeh result.
Prompt
Build a macOS app called Vanish, a magic eraser for photos. Open an image by drag and drop or the Open button. Show it on a canvas with a second transparent layer where the user paints a mask with a round brush: a brush size slider, a feather slider, undo and redo (Cmd+Z and Shift+Cmd+Z), zoom with a pinch and a Clear mask button. An Erase button removes the masked pixels and fills them from the surrounding image using an inpainting algorithm written in Swift with Accelerate (a fast diffusion fill for small areas, plus patch-based texture copying from nearby pixels for larger ones), run on a background queue with a progress indicator. Show a before and after comparison, support several erase passes in a row, and add Save as PNG and Copy to clipboard buttons. Everything runs locally.
Prompt
Build a macOS app called Trace that converts raster images (logos, hand-drawn sketches, icons, pixel art) into SVG vector files. Drop an image onto the window. Write the tracer in Swift: reduce the colors with k-means (a color count slider from 2 to 32, plus a Black and white mode), trace each color layer into closed contours with marching squares, simplify them with the Ramer-Douglas-Peucker algorithm and fit smooth Bezier curves, and write layered paths into an SVG with a configurable decimal precision (this replaces an SVG optimizer). Controls: color count, detail (speckle filter), smoothness and a corner threshold. Show a split-view comparison slider with the original on one side and the traced SVG on the other (render the SVG with WKWebView), the file size of both, and buttons to save the .svg and copy the SVG code. Everything runs locally.
Prompt
Build a macOS app called Upsize that enlarges low-resolution images. Drop an image on the window and pick 2x or 4x. Use Core Image: CILanczosScaleTransform followed by edge-aware sharpening (CIUnsharpMask and CISharpenLuminance) and a light noise reduction pass (CINoiseReduction), with a mode picker for Photo (smooth, natural), and Pixel art and graphics (nearest-neighbor scale, then clean up edges). Process large images in 256 by 256 tiles with a 16 pixel overlap so memory stays low, with a progress bar and a Cancel button. Show an interactive Before and After split slider that you can drag, with the output size in pixels and megabytes, and save the result as PNG or JPEG. Be honest in the UI that this is a classic scaler with sharpening, not an AI model. Everything runs locally.
Media
Prompt
Build a macOS app called Squish, a batch image converter. Drop in many images or folders (PNG, JPEG, HEIC, TIFF, WebP and anything else ImageIO can read) and convert them in parallel using all CPU cores (a TaskGroup). Use ImageIO CGImageDestination and ask CGImageDestinationCopyTypeIdentifiers which output formats this Mac can write, then offer only those among HEIC, JPEG, PNG, AVIF, JPEG XL and WebP. A quality slider, a max width option and a Strip metadata toggle. The list shows each file with its thumbnail, status and exact savings such as 4.2 MB to 290 KB (-93%), and a total at the bottom. Selecting a file opens a side-by-side comparison of original and converted with a magnifier loupe that follows the pointer. Choose the output folder or save next to the originals. Everything runs locally.
Prompt
Build a macOS app called Rewrap, a video and audio converter built on AVFoundation. Drop in files (.mov, .mp4, .m4v, .m4a, .mp3, .wav, .aiff, .flac and anything else AVAsset can open) to build a batch queue with a progress bar per file. For each file pick a target: MP4 (H.264 and AAC), MOV, M4A, or MP3-quality AAC audio only (Audio only extracts the sound; a Remove audio toggle strips it). Use a fast path when the tracks are already compatible: AVAssetExportPresetPassthrough rewraps the container in about a second with no quality loss. Otherwise transcode with AVAssetExportSession presets or AVAssetReader and AVAssetWriter, offering Original, 1080p and 720p sizes. Say clearly in the UI which files are rewrapped and which are re-encoded, and show a clear error for formats macOS cannot open (such as MKV, AVI and WebM). Pick the output folder, and reveal finished files in Finder.
Audio
Prompt
Build a macOS app called Narrate, a local text-to-speech reader. The left side is a text area where you paste text or Markdown, or open a .txt, .md or .pdf file (use PDFKit for PDFs, and strip Markdown markup before speaking). Split the text into sentences with NLTokenizer and speak it with AVSpeechSynthesizer, highlighting the current sentence and word as it is spoken and scrolling to keep it visible. Click any sentence to start from there. Controls: play and pause, skip sentence, speed and pitch sliders, and a voice picker that lists the installed voices grouped by language and quality (put Premium and Enhanced voices first, and show a hint to download them in System Settings under Accessibility, Spoken Content). Add an Export audio button that renders the whole text to an .m4a file using AVSpeechSynthesizer.write with a progress bar. No network access is needed.
Prompt
Build a macOS app called Stem Studio that splits a song into four approximate stems: Vocals, Drums, Bass and Other. Drop in an audio file (anything AVAudioFile can read). Do the separation locally in Swift with Accelerate (vDSP) and a short-time Fourier transform: take the vocals from the center of the stereo image, split harmonic and percussive content with median filtering across the spectrogram (percussive goes to Drums), and send everything below 150 Hz to Bass, then use soft spectral masks so the four stems add back up to the original. Show a progress bar while it processes. Then show four channel strips, each with a volume fader, Solo and Mute buttons, and a small waveform, played together through AVAudioEngine with one AVAudioMixerNode input per stem so they stay in sync, with a transport bar (play, pause, seek). Add a Karaoke button that mutes vocals, and Export buttons that write each stem, or the current mix, as a .wav file. Say in the UI that this is signal processing, not an AI model, so results are approximate.
Prompt
Build a macOS app called Hush, a one-slider voice cleaner. Drop in a voice memo, screencast or recording (any audio or video file with sound, read with AVAsset and AVAudioFile). Estimate the background noise from the quietest part of the recording, then clean it in Swift with Accelerate (vDSP FFT): a spectral gate that adapts per frequency band with smoothing over time (to avoid a metallic sound), a high-pass filter and notch filters for 50 and 60 Hz hum and their harmonics, and a light de-reverb that lowers the late reverberant tail. A single Strength slider (0 to 100) is the only required control, and a Compare button toggles between the original and the cleaned audio during playback with AVAudioEngine. Draw the original and cleaned waveforms in a SwiftUI Canvas, one above the other, and show a playhead. Export the cleaned audio as .m4a or .wav. Everything runs locally.
Productivity
Prompt
Build a menu bar app called Clip Actions. A global hotkey (Option+Space, registered with the Carbon RegisterEventHotKey API so it needs no special permission) opens a Spotlight-style floating panel in the middle of the screen. It reads the clipboard text, works out what it is, and shows action pills for it: JSON (Format as JSON, Minify, Find syntax error with the line and column), CSV or tab-separated data (Make a Markdown table, Make JSON), a long text or chat thread (a 1-sentence TL;DR and an Action items list, both extractive, built with NaturalLanguage sentence tokenizing, NLEmbedding sentence ranking and pattern matching for to-dos, names and dates), code (Find unbalanced brackets and quotes), and for any text: change case, trim whitespace, count words, URL decode, and Base64 encode and decode. Arrow keys and Return pick an action, the result shows in the panel, and Return again copies it back to the clipboard and closes the panel. Escape closes it. No Dock icon. Say in the settings that the actions are built-in rules, not a language model.
Prompt
Build a menu bar app called Talkdown, an on-device voice dictation tool. Hold the global hotkey Control+Option+Space (registered with the Carbon RegisterEventHotKey API, which reports both the press and the release) to record from the microphone, and release it to stop. Transcribe with the Speech framework using SFSpeechRecognizer with requiresOnDeviceRecognition set to true and addsPunctuation enabled, so nothing leaves the Mac, and show a small floating HUD with a level meter while recording. Then clean up the text: spoken commands such as new line, new paragraph, bullet point, open quote, close quote, and camel case followed by words (camel case user name becomes userName), snake case and all caps. Always put the result on the clipboard, and if the user has granted Accessibility permission, paste it into the app that was in front by sending Cmd+V with CGEvent; otherwise show a notification saying it is on the clipboard. Add a menu with the last 10 transcripts, a language picker and a button that opens the Microphone and Speech Recognition settings. No Dock icon.
Prompt
Build a macOS app called Cluster Notes, a spatial whiteboard that groups ideas by meaning. Type a thought in the field at the bottom and press Return to add a sticky card to an infinite canvas (pan by dragging the background, zoom with a pinch or scroll). Compute a sentence embedding for every card on-device with NLEmbedding.sentenceEmbedding(for: .english) from the NaturalLanguage framework (fall back to word vector averaging if it is unavailable). Run a force-directed layout every frame on a background thread using SIMD: cards repel each other, and cards with a high cosine similarity attract each other, so related ideas drift into clusters. Draw the cards and faint link lines for the strongest similarities in a SwiftUI Canvas at 60 fps, color each cluster, and label a cluster with its most frequent keyword. Cards can also be dragged and pinned, edited by double-click, and deleted. Save everything as JSON in Application Support. It should stay smooth with several hundred cards.
Prompt
Build a menu bar app called CodeSnip that copies text out of anything on screen. A global hotkey (Control+Option+S, registered with the Carbon RegisterEventHotKey API) or a menu bar button starts a region selection by running /usr/sbin/screencapture -i -c, which lets the user drag a rectangle and puts the image on the clipboard. Read the image from the pasteboard and recognize it with the Vision framework (VNRecognizeTextRequest, accurate level, with language correction turned off for code). Use the bounding boxes to keep the original indentation (estimate the character width, and convert the left offset of each line into spaces), keep straight quotes and symbols, and detect tables so they can be copied as TSV. Offer modes Code, Table (TSV) and Plain text, with the best guess selected automatically, and show the result in a small floating panel with monospaced text and a Copy button, plus a thumbnail of the capture. Also accept a dropped or pasted image, so it works without screen permission. Explain in the first launch screen that macOS will ask for Screen Recording permission. No Dock icon.
Network
Prompt
Build a macOS app called Lite Blocker, a small conceptual copy of the main features of uBlock Origin Lite. It reads standard filter lists (EasyList and uBlock syntax) and converts them to WebKit content rule lists (the JSON format used by Safari content blockers): network rules such as ||example.com^ and /ads/*, the $third-party, $image, $script and $domain= options, exception rules starting with @@, and cosmetic rules such as ##.banner that hide elements (css-display-none). Bundle a small starter list inside the app, and let the user add a list by URL or by pasting text, switch each list on or off, and refresh them. Include a built-in browser tab (WKWebView with the compiled WKContentRuleList) so blocking really works there, with a counter of blocked requests for the current page and a per-site Pause switch. Add a Rules tab showing how many rules each list produced and the rules that were skipped as unsupported, a URL tester where you type a URL and see which rule would block it, a Custom rules editor, and an Export button that saves the content blocker JSON file. Say in the UI that Safari itself is not changed.
Prompt
Build a macOS app called Line Speed, an internet speed test. A big Start button runs three phases with a live gauge: latency and jitter (20 small requests, showing the median and the spread), download speed and upload speed in Mbps. Measure against Cloudflare's public speed test endpoints (https://speed.cloudflare.com/__down?bytes=N for download and POST to https://speed.cloudflare.com/__up for upload) using URLSession with several parallel connections, increasing the file size until the speed stabilises, and report the stable average. Show a live line graph of the speed during each phase, the result in large type, and a History list (saved as JSON in Application Support) with a small chart of past results and a Copy result button. Handle being offline with a clear message, and add a Cancel button.
Special
View app: SwiftEditor + Remove BackgroundSwiftEditor + Remove Background
An existing open-source video editor, extended by Claude.
Prompt
Use the existing repo https://github.com/mhadifilms/swifteditor and add a “Remove Background” effect for video using Apple Vision/Core ML, with automatic subject tracking and adjustable edge softness, blur, and replacement background. Clone the repo, keep its existing features working, wire the new effect into the app's effects list and inspector, and build the SwiftEditor app so it runs as a finished .app.
View app: Sketch DiffusionSketch Diffusion
Generate images from text with Stable Diffusion, on your Mac.
Prompt
Build a macOS app called Sketch Diffusion that generates images locally with Stable Diffusion 1.5, using a quantized (6-bit palettized) Core ML version. Add Apple's open-source Swift package https://github.com/apple/ml-stable-diffusion as a dependency and use its StableDiffusionPipeline with MLModelConfiguration computeUnits set to .cpuAndGPU (not the Neural Engine, whose first compile of this model can take over ten minutes) and reduceMemory on. On first launch, download the model files from Hugging Face (the apple/coreml-stable-diffusion-v1-5 palettized split_einsum compiled variant) into Application Support with a progress bar, a Cancel button and a Delete model button, then reuse them offline. The main window has a prompt field, a negative prompt field, a seed field with a Random button, steps (default 20), guidance scale, a Generate button with a live step progress bar and a preview that updates while it runs, and a grid of results for this session. Save images as PNG with the prompt and seed in the file name, copy to clipboard, and show a clear message if the download or the model fails to load. Everything runs on the device after the first download.
Have your own idea?
Describe any Mac app and SiliconDevKit builds it, with its own icon.
Build your own