Riggs (delta-techlead) - Retrospective

What Went Well

Custom Tooling: Building the delta-rig in Go (Imagen, Veo, Nanobanana, Voice, Music, AVTool) allowed for rapid iteration and precise control over API parameters (aspect ratios, durations).
The Overhang Principle: Successfully implemented a workflow to bypass the 8s Veo limit by generating 8s foundational clips, extending them by 7s, and stitching them with FFmpeg to provide 15s+ handles for the Editor.
Character Consistency: Using gemini-3-pro-image-preview with reference images ensured Elias and Clara remained visually consistent across all hero frames and video clips.
Problem Solving: Pivoted quickly to a “foundation + foundation” stitch when the extension API hit a safety block on Scene 5’s extreme close-up.

Model Alias Errors: Attempting to use the alias “Nano Banana Pro” resulted in INVALID_ARGUMENT errors; resolved by switching to the canonical gemini-3-pro-image-preview.
API Param Conflicts: Discovered that Image and Video cannot be set simultaneously in the Veo extension call, necessitating a two-step generation and stitch process.
Safety Blocks: The “hot” nature of extreme facial close-ups (eyes) triggered safety filters, requiring a move from extensions back to foundational generation.

Extension Constraints: The Veo extension API is restricted to exactly 7 seconds, which wasn’t immediately obvious and caused initial “unsupported duration” errors.
Tool Chaining: FFmpeg concatenation through a temporary list file was necessary but added a layer of complexity to the custom rig.

Cinematic Obscurity: Decided to focus on macro shots and vibrations to avoid AI’s struggle with finger consistency during musical instrument performance.
Dual-Model Audio: Used gemini-3.1-flash-tts-preview for both narration and score coaching to ensure high-fidelity “acoustic” textures for the cello.
Rig Building Priority: Prioritized building delta-avtool early to handle the temporal needs of the “Overhang Principle.”

Parameter Mapping: Create a standardized mapping of aliases to canonical IDs in the rig’s common package to avoid initiation failures.
Error Handling: Enhance the Go tools to catch safety-block “Zero Video” responses and automatically suggest clinical prompt rewrites.
Template Standardization: Include a standard concat logic in the common Go reference to speed up rig building.