Better Sora Prompts: The Art of AI Direction

I wanted to show the gap between what most people type into Sora and what actually gets great results, because the difference is wild. I started with a simple prompt — “hastily taken selfie by a cat in front of Wrigley Field” — and then ran that same idea through ChatGPT to rewrite it with more descriptive detail. Same concept, same intent, but the enhanced version had specific visual language: lighting direction, lens characteristics, texture descriptions, compositional notes. The output quality jumped dramatically. The first version gave you “a cat at Wrigley Field.” The rewritten version gave you something that actually looked and felt like the reference image style I was going for.

Using custom GPTs for prompting

I also introduced custom GPTs — specifically a Sora Video Prompter — that are trained to write better Sora prompts from the start. Instead of guessing at what descriptive language Sora responds to, you can lean on a purpose-built tool that already understands the model’s strengths. It’s a layer of prompt engineering that most people skip because it feels like extra work, but it’s honestly the highest-leverage thing you can do if you’re serious about getting consistent quality from these generators.

The human layer

Here’s the thing I keep coming back to though: don’t just copy-paste what the AI gives you. Read it. Adjust it. Be a human in the loop. The custom GPT or ChatGPT rewrite gets you 80% of the way there, but that last 20% — the part where you tweak the tone, swap out a word that doesn’t feel right, push the composition in a direction that matches your actual vision — that’s still your job. The tool is a collaborator, not a replacement. The best results I’ve gotten from Sora have always involved that back-and-forth: AI drafts, human refines, AI executes.