Cinematic Production Agent

An autonomous AI agent that turns a one-line episode idea into a full animated production package prompts, dialogue, shots, and captions with Gemini, ADK, and Firestore.

Comment

Inspiration

I run Ajuflow and produce an AI-animated series called the Amara Series (Faith Cinematic niche). Every episode takes hours of manual work: writing character-consistent scene prompts, dialogue, shot direction, then a hook, caption, and hashtags for posting — all tracked by memory with no system keeping continuity between episodes. When this hackathon's Taskmaster track asked for a real, messy multi-step chore to automate, this was the obvious one.

What it does

Given a one-line episode brief, the Cinematic Production Agent autonomously:

  1. Checks Firestore for an existing character bible for the series.
  2. If none exists, invents one (appearance, personality, world rules) and saves it.
  3. Breaks the episode into 4–6 scenes.
  4. Writes a cinematic generation prompt, dialogue, and shot direction for each scene.
  5. Writes a hook, caption, and hashtags for posting.
  6. Saves the full episode package back to Firestore, building a persistent production record over time.

One input, a full production-ready package out — no back-and-forth required.

How I built it

  • Model: Gemini 3.5 Flash via the Gemini API (Google AI Studio)
  • Agent framework: Google ADK, orchestrating tool calls and multi-step planning
  • Cloud infrastructure: Firestore (via Firebase), used as persistent agent memory for character bibles and episode history

Challenges I ran into

Google Cloud's postpaid billing isn't available in Nigeria, and activating a paid billing account required a prepayment I couldn't afford. I solved this by using Google AI Studio's free-tier Gemini API key (no billing account needed) and Firebase's free Firestore tier instead of Cloud Run — proving a fully functional, genuinely cloud-backed agent is possible without any billing setup at all.

What I learned

Letting the agent build its own character bible when none exists — rather than requiring pre-structured input — made the workflow genuinely autonomous instead of a fill-in-the-blanks generator. Firestore doubling as both configuration and growing history means the agent's consistency should improve automatically as more episodes are produced.

What's next

Seeding the agent with my real Amara Series canon (rather than letting it invent its own), and connecting its output directly into my Google Flow video generation pipeline.

Built With

Share this project:

Updates

Submission history