AI Skill Report Card
Refining AI Video Generation Prompts
Quick Start13 / 15
When a user gives feedback like the message above, don't generate the asset yourself — produce a refined prompt/spec that a generation model (image/video AI) or a production pipeline can use. Structure the response as:
- Keep (explicitly preserved elements)
- Change (elements to remove/replace, with concrete substitutes)
- Add (new details requested, made specific)
- Realism polish pass (cross-cutting quality bar applied to every element)
Example output for the input above:
KEEP:
- Mask design/style and orientation exactly as previous version (same mask geometry, angle, and placement on face)
CHANGE:
- Remove the metallic/iron frame or rig around the mask — character's head, neck, and shoulders should read as a real human body (natural skin texture, realistic hair, natural neck movement), with the mask sitting on the face the way a real mask would (visible edges, slight skin-mask contact shadow, subtle fabric/strap if applicable)
- Typing motion: replace mechanical, uniform keystrokes with human-like typing — variable timing between keystrokes, natural finger curl and wrist movement, occasional micro-pauses, slight body sway/breathing while typing, fingers making realistic contact with keys (slight give/bounce)
ADD:
- Background: smoke effect — thin, drifting, semi-transparent, affected by ambient light (not a flat overlay; should have depth/parallax and dissipate naturally)
- Background: ATM/cash machine — realistic materials (brushed metal, LED screen glow, subtle screen flicker), correctly scaled and positioned in depth relative to character
REALISM PASS (apply to all elements):
- Consistent light source and shadow direction across character, mask, smoke, and ATM
- Skin/fabric/metal textures with micro-detail (pores, fabric weave, metal scratches) — avoid CGI-smooth surfaces
- Natural motion blur on fast movements (typing, smoke drift)
- Depth of field: sharp focus on character/mask, slight background softness for the ATM to sell depth
- No floating/disconnected elements — every object should have grounded contact shadows
Recommendation▾
Include a second example with a different domain (e.g., image-only edit, or a case with minimal changes) to demonstrate range beyond the single mask/ATM scenario
Workflow13 / 15
Progress:
- Step 1: Identify what the user explicitly wants kept (don't touch it, restate it precisely so it isn't lost in revision)
- Step 2: Identify what's disliked and translate vague complaints ("looks like a robot," "I don't like this iron thing") into concrete visual/technical fixes
- Step 3: Identify new additions requested (background effects, props) and specify material, physics, and placement details
- Step 4: Apply a global realism pass — lighting consistency, texture detail, motion naturalness, depth/shadow grounding — across old and new elements alike
- Step 5: Output as a structured prompt/brief (Keep / Change / Add / Polish), not vague prose, so it's directly usable for regeneration
Recommendation▾
Add a brief note on handling ambiguous feedback where 'keep' and 'change' conflict, or when user requests are contradictory
Best Practices
- Always separate "what stays" from "what changes" first — prevents regressions where a fix breaks a part the user liked.
- Convert subjective complaints into objective, checkable specs. "More realistic" → specific texture/motion/lighting terms.
- "Robotic" almost always means: uniform timing, no variation, no physics/weight. Fix = add variability, blur, micro-imperfection.
- When user says "every detail should be perfect," treat it as an instruction to add a realism-pass section covering lighting, texture, shadow, depth, and motion — not just the explicitly mentioned objects.
- Name material properties (metal, fabric, skin, smoke) explicitly — generation models respond much better to material/physics language than to "realistic" alone.
- Preserve continuity: same mask, same angle, same framing — only vary what was asked.
Common Pitfalls
- Don't rewrite the entire scene from scratch — the user liked the mask direction; only touch what was criticized or requested.
- Don't leave "make it realistic" unspecified — always cash it out into concrete cues (texture, lighting, motion, physics).
- Don't add background effects as flat overlays — specify depth, transparency, light interaction, and grounding shadows so they integrate with the scene.
- Don't ignore implicit requests — "כל פרט יהיה מושלם" (every detail should be perfect) is a request for a full realism pass, not just decoration.
- Don't output vague creative language back to the user when the goal is a usable generation prompt — always structure and specify.