피파 한 줄 정리: Prompt는 세 가지 방식으로 깨져: under-specified (default soup), over-specified (confused model), conflicting (tug-of-war). 진단부터 하고 고쳐야 해.
Prompts fail in three characteristic ways, and learning to diagnose which one you're hitting will save you enormous amounts of time. Think of it like Goldilocks: too little direction, too much direction, or contradictory direction.
Underspecification: The "Default Soup" Problem
When your prompt is too vague, the model fills in all the unspecified details with its defaults — which means you get generic, conventional, "stock photo" results.
"a woman in a city"
"A woman in her 40s in a dark green trench coat crossing a rain-slicked Tokyo intersection at night, neon reflections on wet pavement, candid street photography, Leica M10, 28mm lens"
The underspecified version will give you: a conventionally attractive woman, mid-20s, in a generic city, probably New York or a stock-photo-looking urban scene, well-lit, centered, and boring. Every unspecified detail defaults to "most common pattern in training data."
Overspecification: The "Confused Model" Problem
The opposite extreme: cramming so many requirements into a prompt that the model can't satisfy all of them and produces a muddled, flat, or oddly composed result.
"A tall thin elderly man with a white beard wearing a blue plaid flannel shirt and worn brown leather boots standing next to a red 1965 Ford Mustang convertible in front of a yellow farmhouse with a wrap-around porch and a white picket fence with exactly four sunflowers to his left and a golden retriever sitting at his feet and a blue sky with three cumulus clouds and a red barn visible behind the house on the right side"
"An elderly farmer in plaid flannel standing beside a classic red Mustang convertible, yellow farmhouse and red barn behind him, golden retriever at his feet, sunny rural America, Norman Rockwell vibes"
The overspecified prompt tries to control every single element, including exact positions and counts. The model will satisfy some of these demands while mangling others. The focused version communicates the same feeling while giving the model room to compose the scene naturally.
Prompt Conflict: The "Tug-of-War" Problem
This is the subtlest failure mode: your prompt contains contradictory or incompatible directions that pull the model in opposing directions.
"Photorealistic anime oil painting in watercolor style, digital art, pencil sketch"
"Detailed watercolor illustration with soft blending and visible paper texture"
"A cheerful, dark, moody, bright, grim, colorful scene"
"A bittersweet autumn scene — warm golden light but bare, lonely trees"
"A person standing underwater in a desert, on fire, covered in snow"
"A surreal scene: a person standing in a desert as snow falls from a clear blue sky, dreamlike atmosphere, Magritte-inspired"
Diagnostic Flowchart
Output looks generic/boring? → Probably UNDERSPECIFIED → Fix: Add subject, lighting, composition, style specifics Output looks confused/muddled? → Probably OVERSPECIFIED → Fix: Remove less-important constraints, focus on 3-4 key elements Output looks weird/inconsistent? → Probably CONFLICTING → Fix: Check for contradictory style/mood/physics terms
- Underspecified: Too vague → generic defaults. Fix by adding specific details for subject, lighting, and style.
- Overspecified: Too many demands → confused, muddled output. Fix by focusing on 3-4 key elements.
- Conflicting: Contradictory terms → weird, averaged results. Fix by ensuring stylistic and physical coherence.
- Most prompt problems fall into one of these three categories. Learn to diagnose which one you're hitting.