Screen Studio-style screen recorder app for Mac and Windows

[Note: the creator shared a partial prompt. Wording below is theirs.] GOAL: You're building Dolly, a screen recorder for macOS and Windows. MOTION: It does what Screen Studio does, rebuilt from the screenshots I attached. Don't copy its code, name, icons or assets. What it does: record the screen, then produce a polished video automatically: zoom on every click, smooth cursor, wallpaper background with padding, webcam bubble, on-device captions, export. Stack: one Electron + React + TypeScript codebase. FFmpeg (ffmpeg-static) for encoding, whisper.cpp for captions. Business model: open core. - Free (MIT, forever): unlimited recordings, auto-zoom, smooth cursor, backgrounds and padding, MP4 export up to 1080p 30 fps, no watermark. - Pro ({{PRICE}} one time): 4K and 60 fps export, captions, webcam bubble, GIF export, 9:16 vertical cut, custom background images. - Payments: Whop license keys, sold at {{CHECKOUT_URL}}. Read Whop's current API docs for validating a license key and binding it to a device. - Never ship the Whop API key inside the app. Add a tiny Vercel serverless function (/api/validate) that takes the key + a device ID, calls Whop, and returns valid / invalid. The API key lives in a Vercel env variable. - Settings has an "Enter license key" field. Validate once online, cache the result so Pro keeps working offline. Allow 3 devices per key. - Locked Pro features stay visible with a small "Pro" badge. Clicking one opens a short upgrade sheet that links to {{CHECKOUT_URL}}. Never block or nag free features. Process: 1. Before any code, design 6 concepts for the editor (preview canvas, timeline with zoom blocks, export panel, upgrade sheet) as static HTML in /concepts. Stop and wait for me to pick one. 2. Write SPEC.md: features, free vs Pro split, file layout, the project data model (clips, cursor events, zoom segments) and acceptance checks. Wait for my OK. 3. Split the build across parallel subagents, one per area: - capture: screen + mic + webcam, log every cursor move and click - zoom engine: one zoom per click, merge clicks under 1.2 s apart, 2.0x default, 0.6 s ease in/out, never show past the frame edge - cursor smoothing: spring, not linear - editor UI + timeline: drag, stretch and delete zoom blocks - export: MP4 H.264 and GIF, 1080p/4K, 30/60 fps, 16:9, 1:1, 9:16 - captions: whisper.cpp, fully offline - licensing: key check via /api/validate, offline cache, upgrade sheet - installers 4. Run a separate review agent on capture, file-writing and licensing code: macOS Screen Recording + Accessibility permissions, Windows Graphics Capture, where files get written, no secrets in the app, nothing leaving the machine except the license check. Fix everything it finds. 5. GitHub Actions: build the .dmg on macos-latest and the .exe on windows-latest for every tag, attach both to the release. 6. Test it for real: record 20 seconds on this machine, export, check every zoom lands on a click. Test that a fake key is rejected and Pro stays