Showcase: I built a series of connected scenes. It took 47 attempts.

I've been conducting a systematic evaluation of Sora's capabilities for generating consistent, multi-shot narratives for potential use in technical training and architecture visualization. The objective was to produce a sequence of four connected scenes featuring the same protagonist (a woman with specific, identifiable clothing and features) in a consistent cyberpunk-style environment. The narrative arc was simple: entering a data hub, interacting with a holographic terminal, receiving an alert, and reacting.

The results were quantitatively poor but qualitatively illuminating. Out of 50 total prompt attempts, only 3 generated clips could be considered to have maintained acceptable character and environmental consistency when sequenced. The failure modes were highly predictable and fell into distinct categories:

* **Character Inconsistency:** This was the most frequent failure. The protagonist's hairstyle, hair color, facial structure, and jacket design would change drastically between shots, even with extremely detailed prompt anchoring. Subtle elements like a specific emblem on the jacket appeared in maybe 1 out of 10 generations.
* **Environmental Drift:** The layout of the "neon-lit data hub" would reconfigure itself entirely. A central terminal in one shot would be against a wall in the next. Lighting color schemes (e.g., "dominant cyan and magenta accent lighting") proved to be a moderately stable anchor, more so than physical geometry.
* **Action Continuity:** The flow of motion was often discontinuous. A character reaching for a console in clip A would be shown already touching it in clip B from the same angle, breaking the natural progression.

My most successful approach involved a form of "prompt engineering through iterative constraint," building a base prompt and locking down variables. The final, moderately successful prompt sequence looked like this:

**Base Anchor Prompt:** "Cyberpunk data hub interior, neon cyan ambient light with magenta accents. A woman, early 30s, with a precise chin-length black bob haircut and a silver tech-weave jacket with a triangular shoulder emblem, stands before a large holographic interface. Cinematic, 35mm, detailed."

**Scene-Specific Overrides:**
* Scene 1: `[Base Anchor Prompt], wide shot, she enters through a sliding door, looking around the hub.`
* Scene 2: `[Base Anchor Prompt], medium shot, her fingers manipulate glowing holographic data panes.`
* Scene 3: `[Base Anchor Prompt], close-up on her face and the hologram, which suddenly flashes red warning glyphs.`
* Scene 4: `[Base Anchor Prompt], low-angle shot, she turns sharply, looking off-screen, her hand moving to an earpiece.`

Even with this structure, achieving the 3 usable clips required generating 5-7 variants for *each* scene prompt and selecting the ones with the highest mutual consistency. This is not a workflow for rapid iteration. The cost in time and computational resources (even abstracted through the API) is significant.

From an infrastructure perspective, this exercise highlights the current limitations of the model as a tool for producing deterministic, multi-asset outputs. It functions as a phenomenal idea generator and single-scene producer, but for any form of consistent storytelling, it requires a human-in-the-loop curation and editing pipeline that is far more extensive than traditional CGI or even game engine workflows. The "47 attempts" in the title is not an exaggeration; it's a sobering metric for anyone considering this technology for projects requiring narrative continuity.