fal.ai product builder — two types, two flows

fal.ai is a generation API and nothing else: no hosting, no end-user accounts, no CLI that owns your project. You generate media with fal and write the code on the local filesystem with git and the repo's package manager. Every build lands as a working local build in the user's repo — the user owns deploy.

The two types

Resolve the type first. The test is one question: who triggers a generation?

• site — only you do, before ship. Stills, film, and cover are generated now, landed into public/generated/, and committed. The shipped bundle carries no fal call and no FAL_KEY; the site renders fine with fal offline. Every site gets an independent brand — palette, type, and chrome from its own design brief. Flow: references/site-flow.md.
• app — the visitor does, on the user's fal bill. Someone types a prompt or drops an image and gets media back, through the fal queue behind a server route that holds FAL_KEY. Flow: references/app-flow.md.

Quick tells: "landing page / portfolio / marketing site / rebuild with no photography" → site. "users generate headshots / thumbnails / voiceovers / clips", "a generate button" → app. A site that wants a generate button is an app; build it as one rather than putting a key in the browser.

An app also generates its own brand assets at build time (shell art, cover), so it reads references/generation.md too. A site never reads the runtime reference.

Intake — one batch, before any generation

Ask everything you still need in a single message, then build: