AI Video Production System
Implementation guide for a UGC-style AI video stack, with JSON prompt schemas
An X long-form article documenting an eight-month build-out of an AI video pipeline for hyper-realistic UGC content. It maps which model to use per shot type, starting with Higgsfield and Nano Banana 2 for character generation, and argues that JSON-structured prompts beat prose prompts because they pin down colour grading, lighting and styling instead of leaving them to the model. It includes a full copy-ready JSON prompt schema with fields for composition, framing, camera height and angle, depth of field, negative space and subject description.
The transferable idea here is the JSON prompt schema, not the specific tool list — lift the field structure (composition, framing, camera_height, camera_angle, depth_of_field, negative_space, subject) and reuse it against whatever image or video model the user actually has access to. Reach for it when generated portraits look plastic or the colour grading collapses to four or five muddy tones, which is the exact failure it diagnoses. Treat the named models as dated on arrival: this space turns over every few months, so verify Higgsfield and Nano Banana 2 are still the right picks before recommending them.
Why it is here. Most AI-video content is vibes; this one publishes an actual prompt schema you can paste and adapt, which is the rare part that survives the next model release.
- JSON prompt schema over prose prompts
- model-per-shot-type mapping
- diagnoses the plastic-skin failure
- copy-ready portrait prompt structure
Verified 2026-08-15. Not fetched per instructions (x.com); record built from the pre-retrieved article title and full body text supplied in the batch.
