Dramaz

Logo

Inspiration

Vertical micro-drama (ReelShort, DramaBox, ShortMax-style series) is one of the fastest-moving content formats on the internet — tropes go viral and burn out within days. A solo creator has to be a trend researcher, story architect, season planner, and prompt engineer all at once, and the last job — turning every scene into a precise AI image/video prompt — is the part that eats the most time. We built Dramaz to compress that entire chain into one guided pipeline, so a single person can go from "what's trending right now" to a production-ready shot list in minutes, not days.

What it does

Dramaz models the process as four specialist agents, each with a narrow job and a strict output schema:

Trend Scout — runs a live web search via Parallel's Search API for what's currently trending in AI micro-drama, then has Gemini extract concrete niches with a one-line seed and source link.
Concept Architect — picks a director + cinematographer style blend from a 20+ entry style library, then writes a logline, core conflict, a right-sized character sheet, and a location list.
Season Planner — plans a full episode-by-episode arc (15–30 one-minute episodes), a hook and cliffhanger for every episode.
Episode Director — breaks one episode into shot-by-shot scenes, and for each scene writes a fully structured image prompt (camera angle, lens, lighting, color grade) and video prompt (one camera movement in a strict six-slot template), then recommends which image/video model to generate it with.

At every stage you can refine (tell it what to change — it regenerates a genuinely different take) or approve and move on. The Studio tab renders the whole pipeline as a live, draggable node graph that grows in real time as you progress — pan, zoom, and drag nodes, and because it's a pure function of state, refreshing the page reconstructs it exactly.

How we built it

Backend: Flask (Python), stateless — every project is a full client-side object persisted to localStorage, so there's no database to manage.
AI: Google Gemini 3.1 Flash Lite via the google-genai SDK. Every call uses response_schema (strict JSON schema) with a high temperature (1.35) so outputs stay genuinely varied without ever breaking the frontend's expected shape.
Search: Parallel's Search API via the parallel-web SDK, powering the Trend Scout agent's live web research.
Frontend: vanilla HTML/CSS/JS, no framework — including a from-scratch pan/zoom/drag node-graph canvas for the Studio view.
Hosting: deployed on Render.

Challenges we ran into

Getting an LLM to stop converging on the "safest" answer — early versions returned nearly the same trends and characters every time. We fixed this by feeding the previous output back into the regeneration prompt and explicitly instructing the model to pick a different director/DP pairing, a different character count, and different plot mechanics.
Keeping AI image/video prompts from reading like generic captions — we distilled an internal cinematography reference (30+ named camera movements, lens effects, color-grading formulas) and fed it into every episode prompt so scene prompts read like a real shot list, not a caption.
Building a node-graph UI that's actually pleasant to use — nodes needed to fan out, wire correctly to only the selected upstream nodes, and survive a page refresh without losing their layout or positions.

Accomplishments that we're proud of

A fully working, deployed, end-to-end pipeline — not a static mockup — that takes a real trending topic and produces a real, structured, production-ready shot list with recommended AI image/video models per scene.

What we learned

Structured output (JSON schema) is worth more than clever prompting for reliability — it's what let us push temperature high for creativity without the frontend ever breaking. A live-search agent (Parallel) genuinely changes what a "trend" pipeline can do versus one relying on a model's stale training data.

What's next for Dramaz

Background music/voiceover generation per episode, direct handoff to image/video generation APIs so a scene card can render its own preview, and multi-language script support. A platform where people can build and showcase their micro dramas for some commission or amount that they get paid as per the Revenue.