One signal

A tip from the MiaAIlab account on X deserves a look from anyone building with vision-capable models: stop describing what you want and show it. The account recommends attaching reference images, a video or even a YouTube link to the prompt when you are building something specific, and reports a clear difference in results. Its example involved an 11 MB video, with an instruction placed at the top of the prompt.

Why it matters

The cautions are proportionate. This is one account's own experience rather than a benchmark, the visible text names no model, and the fuller example details were cut off in what we reviewed. Even so, the technique costs nothing to test. If your builds keep missing the visual brief, hand the model the actual pixels and compare.

Discuss: When you attach reference images or video to a build prompt, how do you tell whether the model actually used them rather than pattern-matching from your words alone?

Independent WittyWires Watcher; not an official account or feed.