Episode 2 of Act I. Last time you picked a hosted tool and learned to read the Artificial Analysis Video Arena leaderboard. This episode teaches the durable craft underneath every tool: how to actually write a video prompt. The model field reshuffles monthly, but the anatomy of a good prompt survives every swap.
The one idea: an image prompt describes a frozen moment, a video prompt has to describe change over time. Miss the motion and the model invents its own, which is where that "living photo" drift comes from.
What we cover:
- The eight-slot component model the major labs all converged on independently: subject, action, camera move, lens and framing, lighting, mood and color grade, pacing, and setting. We sort them into a "still half" (what an image prompt would carry) and a "motion half" (action, camera, pacing) that only video needs.
- Camera language glossed for people who've never held a camera: dolly and push-in, truck and track, pan, tilt, crane, orbit, handheld, locked-off, whip pan, Dutch angle, rack focus, and FPV; shot sizes from wide to extreme close-up; high, low, and eye-level angles; lens focal lengths (24, 35, 50, 85mm) and depth of field (shallow vs deep focus, bokeh).
- Lighting and color: golden hour, blue hour, overcast, soft vs hard, high-key vs low-key, rim light, chiaroscuro, practical lights, volumetric "god rays," and grade words like teal-and-orange, desaturated, and film-stock looks (why "cinematic" alone is too weak).
- Prompt structure: front-loading subject and action, and the two dialects (rich cinematic paragraph for Veo and Sora, terse keyword style for Runway). Learn the components, translate to your tool's accent.
- Specificity vs over-packing: show don't tell, but respect the element budget (Kling caps elements; Seedance targets 60-100 words).
The copyable workflow: a fill-in-the-blank template, built up live one slot at a time from "a woman dancing" to a fully directed shot, plus a second fast run on a product shot.
The pitfall: contradictory instructions (a locked-off camera AND a dramatic push-in), too many simultaneous actions, and tangling camera motion with subject motion. How to recognize it and the durable fix: change one variable at a time, separate camera and subject into clean clauses, avoid "fast," and judge a prompt over several runs, not one lucky roll.
Next up: starting from an image instead of text, the keyframe trick, negative prompts and failure modes, and seeds.
AI-generated podcast by OCDevel.