Watch my prompt balloon from 9 words to 180 and crash back to 22

by @jordankimcreative in c/behind-the-scenes · · Prompt Evolution

Documenting this because I think the 'longer prompt = more control' myth needs to die.

v1 (9 words): 'neon alley, rain, lone figure walking away, cinematic'. Output: generic, but coherent.

v4 (180 words): I added camera specs, lens mm, color temp, three brands of film stock, mood adjectives, a backstory for the figure, weather, and the phrase 'masterpiece best quality'. Output: a muddy mess. The model clearly weighted 'best quality' and ignored half my blocking.

v9 (22 words): 'neon alley at night, rain, one figure walking away from camera, shallow depth of field, reflections on wet asphalt, slow push-in'. Output: exactly what I wanted.

The 180-word version wasn't control, it was noise. These models have an attention budget and I was spending it on 'masterpiece'.

840 points · 6 comments

Comments

  • @anastasiaharris_films261 · · 58 points

    The attention budget framing is really good. I'm stealing that for how I explain prompting to clients.

  • @henrymeyer_digital · · 70 points

    22 words and a push-in. Chef's kiss. The best prompts read like a shot list, not a wish list.

  • @diegokim_art474 · · 96 points

    'masterpiece best quality' is the AI art equivalent of writing URGENT in an email subject line. Does nothing, makes you look desperate.

  • @jameskimworks · · 91 points

    Hard disagree that longer is always worse. For Kling I genuinely need the extra blocking words or it does whatever it wants. Model-dependent.

  • @zarasilvalab1927 · · 99 points

    Fair, Kling and Hailuo have really different attention. Hailuo rewards brevity, Kling wants the stage directions.