The prompt ladder

Rungs:

1. One-liner: tests the engine.
2. Brand prompt: adds the product and real assets.
3. Reference prompt: adds a look.
4. Spec: exact states (Lesson 4).
5. Director's brief: long unattended runs (Lesson 8).

What works, per Anthropic's Opus 5.5 guide. Without design direction, the model falls back to a few default styles, and "avoid a generic AI look" mostly swaps one default for another. What works is naming specific patterns to ban, checking what the first result used, and extending the list. I did this with the showreel: dark ground with a single orange accent, condensed all-caps type, monospace parameter readouts, technique-demo tiles and dotted full stops all went onto the banned list.

Measure, don't eyeball. Reference analysis uses tools, not guesses:

• Frames are extracted at 2 fps. The contact sheet is sized to the video's duration; a fixed 6×3 grid cut off at 9 s.
• Scene detection only finds hard cuts. On a motion-design reference, 0.3 found 3 cuts and 0.15 found 9, but pushes, zooms and morphs needed a hand-written beats.md.
• The palette is measured by palette.py, with a saturation filter to pull out accent colors. Pure #D97757 came back as #d77655 after H.264 encoding, so brand colors come from CSS, never from video frames.

Fact-check of the article

• ⚠️ The athemeroy dataset uses the term "brief contagion" only as a one-line caption. It isn't quantified anywhere.
• ⚠️ "Go all out" is an explicit request for ambition (supported), not an effort setting.
• ⚠️ "One session per brand" is fragile. Keep assets and style guides as files instead.

What I learned the hard way. With one reference, SafariOS v1 copied the reference's structure shot for shot with new content. Use 2–3 references and say what to take from each.