AM3 Engine — Cute Kitty for Codex

A character-continuity pipeline that turns one approved identity into a coherent, validated, and installable Codex Pet. Cute Kitty is its reference implementation.

Comment

Inspiration

AI companions are memorable only when their identity and behavior remain coherent. Most image-generation workflows produce attractive frames but lose the character across scenes, states, and channels. We treated character continuity as an engineering contract.

What it does

AM3 Engine transforms one approved character definition into a state-aware, validated, and installable Codex Pet. Cute Kitty is the reference implementation: a curious feline AI companion with a governed visual identity, nine acting scenes, channel-ready media, and a one-click or agent-assisted installation path.

The public showcase lets judges inspect the character, acting scripts, designed durations, production path, and installation package. The same source identity governs the sweater, glasses, laptop, AM3 emblem, performance, and sprite sheet.

How we built it

We used Codex as the primary engineering collaborator to structure requirements, implement the React/Vite portal, prepare Docker publication, reconcile character assets, and package the Codex Pet.

GPT-5.6 supported the high-reasoning work: finding continuity failures, strengthening deterministic validation gates, translating acting intent into timed scenes, and keeping the installation artifact aligned with the approved character.

The showcase is packaged as a standalone Docker image for Azure App Service. The installation flow serves a transparent WebP sprite sheet, metadata, a local installer, and an agent-readable installation prompt.

Challenges

The hardest problem was not generating a single beautiful picture. It was preventing drift across many outputs: face, proportions, costume, emblem geometry, laptop material, acting beats, timings, videos, and the final channel package.

We also had to separate a channel-agnostic character definition from the Codex-specific adapter and installation contract.

Accomplishments

  • Preserved one canonical character identity across the complete delivery.
  • Created nine acting scenes with explicit triggers and designed durations.
  • Prepared a validated and installable Codex Pet package.
  • Added human, web, and agent-assisted installation paths.
  • Built a bilingual and responsive public showcase.
  • Created a reproducible character-to-channel workflow rather than a one-off visual asset.

What we learned

Generative character work becomes more reliable when visual approval, acting direction, timing, and publication metadata are treated as versioned interfaces.

Human approval remains essential at creative gates, while AI agents are especially valuable for traceability, reconstruction, repetitive packaging, validation, and consistency checks.

What's next

We plan to adapt the same character contract to additional digital channels, add more characters without duplicating the pipeline, and turn channel adapters into reusable publishing components.

Built With

Share this project:

Updates