Cloudinary

Promptfree, for everyone, and the only version there is

Copy the prompt for Claude CodeCopy the prompt for CodexCopy the prompt for Cursor

Copy

Build me image and video delivery that replaces Cloudinary — and know why this is ALMOST.

Resizing images is easy. Three things are not: **format negotiation** that serves AVIF, WebP or JPEG based on what each browser actually accepts, **smart cropping** that keeps the subject in frame at any aspect ratio, and **video transcoding** into the ladder of qualities a player needs. The first two are an afternoon each. The third is where the effort and the machine cost really are — read that section before committing.

STACK
- Node 20+ with Fastify
- sharp for images, ffmpeg for video
- SQLite through better-sqlite3 for the asset index and the transform cache
- Local disk or an S3-compatible bucket, with a CDN in front
- Caddy in front

THE DATA MODEL
- assets: id, public_id, kind, original_path, sha256, bytes, width, height, duration_ms, format, colour_profile, focal_x, focal_y, created_at, deleted_at
- derivatives: id, asset_id, transform_hash, params_json, format, path, bytes, width, height, created_at, hits
- video_renditions: id, asset_id, height, bitrate, codec, path, bytes, status
- uploads, folders, tags — the ordinary library shape
- Every derivative is regenerable from the original, so the original is what gets backed up and the derivatives do not

THE URL
- Transformations expressed in the path: width, height, fit mode, quality, format, gravity, and effects
- **Signed.** An unsigned transform URL means anybody can request ten thousand sizes of your images and use your machine as a free rendering farm. Sign the parameters, or maintain an allow-list of named presets — the second is better, because it also stops your own templates from inventing forty variants of the same image
- Cached by transform hash, with immutable headers and a content hash in the path
- Generated once, on first request, then served from the cache forever

FORMAT NEGOTIATION
- Read the browser's Accept header and serve AVIF where accepted, then WebP, then the original format. Vary on Accept so the CDN caches each separately, or the wrong format will be served to somebody
- AVIF is meaningfully smaller and much slower to encode. Encode it lazily and in the background, serving WebP until it is ready
- Quality chosen by a perceptual target rather than a fixed number, measured once per image and cached — quality 80 is wasteful on a flat illustration and visibly poor on foliage

SMART CROPPING
- A focal point stored per asset, set by hand in the interface, is the cheapest and most reliable answer and it covers most cases
- Automatic: face detection, or a saliency heuristic based on edge density and contrast, to propose a focal point that a human confirms
- Never crop automatically without a stored focal point on anything user-facing. A logo cropped in half is worse than a letterboxed image

VIDEO, WHICH IS THE EXPENSIVE PART
- A ladder of renditions — several heights and bitrates — packaged as HLS or DASH so a player can adapt to the connection
- Transcoding is CPU-hours per video. Queue it, cap the concurrency, and be honest about the machine cost. This is the part where the hosted service starts to look reasonable
- A poster frame, and a short preview clip for hover, both generated automatically
- Store the original untouched, always

DELIVERY
- Put a CDN in front. Serving media from one machine is the mistake that makes everything feel slow, and it is the one part of this that money genuinely solves
- Correct cache headers, range support, and a stale-while-revalidate policy on derivatives

OPERATIONS
- .env: DATABASE_PATH, STORAGE_PATH, BASE_URL, SIGNING_SECRET, CDN_URL, TRANSCODE_CONCURRENCY
- Migrations on boot, each once
- Disk watchdog, and a sweeper for unused derivatives by last hit
- Backup the originals; derivatives are regenerable
- Health endpoint that runs one transform through the real path

WHAT MATTERS MOST
Signed or named transforms, and Vary on Accept. Get those two wrong and you have either an open rendering farm or a cache serving AVIF to a browser that cannot read it.