Meeee

byHaji Hasnain

# CreatorSuite AI — Full Build Prompt Copy everything below the line into Claude Code (or your AI builder of choice) as the project brief. --- ## Project Overview Build a SaaS web app called **"CreatorSuite AI"** — an all-in-one content packaging and distribution tool for YouTube/short-form creators. It combines three proven, revenue-validated features into one subscription product: 1. **AI Thumbnail & Title Generator** (validated by Pikzels, ~$25K MRR) 2. **Long-form-to-Shorts Repurposing Engine** (validated by Opus Clip, Vidyo, and similar tools) 3. **Content Scheduler & Cross-Platform Publisher** (validated by Senja-style creator tooling, ~$83K MRR) Target user: solo creators and small teams (1–10 people) who post across YouTube, TikTok, Instagram Reels, and Shorts and are currently stitching together 3–4 separate tools to do this. --- ## Tech Stack - **Frontend**: Next.js 14 (App Router) + TypeScript + Tailwind CSS + shadcn/ui - **Backend**: Node.js API routes (or separate Express/Fastify service for video processing) - **Database**: PostgreSQL (via Supabase or Neon) — users, projects, assets, subscriptions - **File/video storage**: S3-compatible bucket (Cloudflare R2 or AWS S3) - **Auth**: Clerk or Supabase Auth (email + Google OAuth, since creators live in Google's ecosystem) - **Payments**: Stripe (subscriptions + usage-based credits) - **AI services**: - Image generation/editing: fal.ai or Replicate (Flux/SDXL) for thumbnails, with face-swap model - Video transcription: Whisper API (via OpenAI or Deepgram) - Clip selection/highlight detection: Claude API (Sonnet) — feed transcript + timestamps, ask it to identify the N most "hook-worthy" segments with reasoning - Auto-captioning/burned-in subtitles: FFmpeg + Whisper word-level timestamps - **Video processing**: FFmpeg (server-side, run in a queue — not in the request/response cycle) - **Job queue**: BullMQ + Redis (video processing and AI generation are slow — never block the UI) - **Scheduling/publishing**: YouTube Data API, TikTok Content Posting API, Meta Graph API (Instagram/Facebook) --- ## Feature Module 1: AI Thumbnail & Title Generator **Core generation** - Text-prompt-to-thumbnail: user describes the thumbnail, AI returns 4 variations in YouTube's 1280x720 format - Style/Persona training: user uploads 5–10 reference images of themselves; system fine-tunes a lightweight face embedding so their face can be inserted into any generated thumbnail consistently - FaceSwap: upload any generated or existing thumbnail + a selfie → AI swaps the face in convincingly - Thumbnail recreation: paste a YouTube URL → AI extracts and analyzes the thumbnail layout/composition → rebuilds that layout using the user's own subject matter - Sketch-to-thumbnail: rough shape/text layout sketch → polished thumbnail **Title generation** - Input video topic/brief → generate 8–10 title options - Reword / shorten / lengthen / "make it more curiosity-driven" controls on any generated title - Pair titles with thumbnails and let the user preview them together as they'd appear in a YouTube feed **Scoring & iteration (this is the differentiator — don't skip it)** - After generating, score each thumbnail+title pair on 5 axes: Virality, Clarity, Idea strength, Curiosity, Emotion (1–10 each, with a one-line reason per score) - "Fix weak packaging" button: auto-applies improvements based on the lowest-scoring axis - Natural-language editing: "make the text bigger," "change background to red," "make my expression more shocked" — text-in, image-out, no design tools required **Data model**: `thumbnails` (id, user_id, project_id, prompt, image_url, title_text, scores JSON, version_of), `personas` (id, user_id, face_embedding_ref, sample_images[]) --- ## Feature Module 2: Long-form-to-Shorts Repurposing Engine **Ingestion** - User uploads a long-form video (MP4, up to ~3 hours) or pastes a YouTube URL - Backend transcribes with word-level timestamps (Whisper) **Clip detection** - Send the full transcript + timestamps to Claude with a prompt like: *"Identify the 8–12 most self-contained, hook-worthy 20–90 second segments in this transcript. For each, give start/end timestamps, a one-line reason it works as a standalone clip, and a suggested title."* - Display candidate clips as a ranked list with the reasoning shown, not just a black-box output — this builds trust and lets users learn what makes a good clip **Editing** - Auto-crop long-form 16:9 to vertical 9:16, with face-tracking so the speaker stays centered (this is the single hardest technical piece — budget the most engineering time here) - Auto-generate burned-in captions (word-by-word highlight style, like CapCut/Opus Clip) from the Whisper word timestamps - Auto-select a thumbnail-worthy frame for each clip and route it into Module 1's generator for a matching thumbnail - Basic trim/re-time controls so users can adjust the AI's chosen in/out points **Output** - Export each clip as MP4, ready for TikTok/Reels/Shorts dimensions - Batch export all clips from one long-form video at once **Data model**: `source_videos` (id, user_id, url_or_file, transcript JSON, duration), `clips` (id, source_video_id, start_ms, end_ms, caption_style, status, output_url) --- ## Feature Module 3: Scheduler & Cross-Platform Publisher - Connect accounts: YouTube, TikTok, Instagram/Facebook (OAuth per platform — this requires developer app approval from each platform, budget time for this) - Calendar view: drag-and-drop clips/videos onto publish dates and times - Per-platform caption/hashtag fields (don't force one caption across all platforms — this is a common complaint about competitor tools) - Best-time-to-post suggestions based on the connected account's own historical engagement data (pull from each platform's analytics API) - Queue status dashboard: scheduled / published / failed, with retry on failure **Data model**: `scheduled_posts` (id, user_id, clip_id_or_video_id, platform, scheduled_time, status, platform_post_id, caption) --- ## Cross-cutting requirements - **Onboarding**: first-run flow should get a user from signup to their first generated thumbnail in under 2 minutes — this is the activation moment that determines retention - **Credits system**: thumbnails, clip-generation, and AI scoring all consume credits; video export and scheduling do not (keeps the expensive AI calls metered, keeps the "glue" features unlimited so users don't feel nickel-and-dimed on the sticky parts) - **Brand kit**: fonts, color palette, logo, and Persona face — applied automatically across both thumbnail and clip-caption styling so output stays on-brand without re-specifying it each time - All video processing must run as background jobs with progress states surfaced in the UI (queued → processing → ready) — never make the user wait on a blocking request for anything video-related --- ## Pricing (subscription + credits hybrid — mirrors what's already working in this space) | Tier | Price | Thumbnails/mo | Clips/mo | Scheduled posts | Accounts connected | |---|---|---|---|---|---| | Starter | $19/mo | 20 | 10 | Unlimited | 2 | | Creator | $49/mo | 150 | 40 | Unlimited | 5 | | Studio | $99/mo | 450 | 100 | Unlimited | Unlimited | Overage credits sold à la carte so power users aren't capped mid-launch week. --- ## MVP scope (build this first, ship in 4–6 weeks) 1. Auth + Stripe subscription 2. Thumbnail generator (prompt-to-image + title generation + scoring) — no FaceSwap/Persona yet 3. Clip detection + vertical crop + captions for one video at a time — no batch, no scheduler 4. Manual export (download MP4) — no platform publishing yet Add FaceSwap, Persona training, batch processing, and the scheduler/publisher in Phase 2 once you have paying users validating the core loop. The publisher module in particular requires per-platform API approval that takes weeks — start those applications in parallel with MVP development, not after. --- ## Prompt to hand your AI coding tool > Build a Next.js 14 + TypeScript SaaS called CreatorSuite AI. Set up the project with Tailwind, shadcn/ui, Supabase (auth + Postgres), Stripe subscriptions, and Cloudflare R2 for storage. Start with the Thumbnail Generator module: a page where a user enters a text prompt and video topic, calls an image generation API (fal.ai Flux) to return 4 thumbnail varioptions at 1280x720, calls an LLM to generate 8 title options, and displays scored thumbnail+title pairs (Virality/Clarity/Idea/Curiosity/Emotion, 1–10 with a one-line reason each). Include a credits system tied to Stripe subscription tiers (Starter/Creator/Studio). Set up BullMQ + Redis for background job processing since later modules will need it for video work. Use the data models: users, projects, thumbnails, personas — as specified in the schema below. [paste data models from above]