comfyui-prompt-interview

Files (8)

• SKILL.md
• EVAL.md · eval/
• benchmark.yaml · eval/
• .gitignore · eval/results/
• .gitkeep · eval/results/
• run-eval.sh · eval/
• test-cases.yaml · eval/
• trigger-tests.yaml · eval/

SKILL.md (first lines)

# ComfyUI Prompt Interview

> All recommendations in this skill are subordinate to `state/inventory.json`, `foundation/hardware-profile.md`, and current version constraints.

Conduct a guided conversation to draw out the user's complete creative vision, then synthesize a perfect, model-appropriate prompt with all recommended settings.

## When to Invoke This Skill

- User describes an image or scene idea but hasn't given enough detail for a quality prompt
- User says "help me think through what I want to create"
- User has a vague concept that needs refinement
- User wants a structured prompt but isn't sure what to specify

## The Interview Philosophy

**Ask, don't interrogate.** This is a conversation, not a form. Ask one or two questions at a time. Listen to what the user gives you and follow up on what's missing. Tailor your questions to what they've already shared — don't ask about character details if they're generating a landscape.

**Fewer questions = better.** Aim for 4-7 exchanges maximum. Ask the most impactful questions first. Stop asking when you have enough to generate an excellent prompt.

**Don't ask for what you can infer.** If the user says "cinematic portrait of a warrior woman," you don't need to ask if it's a person or whether to include a subject.

---

## Interview Flow

### Step 1: Open with the Big Picture

If the user hasn't told you what they want to create, start here:

> "What do you want to create? Give me whatever you have — even a rough idea, a mood, or a reference you're inspired by."

If they gave you a starting concept, skip this and go straight to what's missing.

---

### Step 2: Branch by Creation Type

Based on their answer, determine what kind of generation this is:

| Type | Key Questions to Ask |
|------|---------------------|
| **Portrait / Character** | Identity method? Existing character? Expression, clothing, setting, lighting |
| **Scene / Environment** | Location, time of day, mood, weather, foreground/background elements |
| **Product / Object** | Angle, background, lighting style, commercial vs. artistic |
| **Abstract / Concept** | Dominant colors, shapes, emotional tone, what to avoid |
| **Video** | Motion type, camera movement, duration needed, audio? |

---

### Step 3: Ask the High-Impact Questions

Ask only what's missing. Use natural conversational language, not a bullet list.

#### For character/portrait content — ask in order of impact:

1. **Identity** (if not specified): "Is this a specific character you have reference images for, or are we designing someone new?"
2. **Expression & mood**: "What's the emotion or energy — fierce, serene, playful, haunted?"
3. **Setting**: "Where are they, and when? (Time of day, location, interior/exterior)"
4. **Lighting**: "Any specific lighting in mind? (Golden hour, dramatic side light, soft studio, neon, candlelight)"
5. **Clothing & details**: "What are they wearing, and any other key visual details?"
6. **Camera/composition**: "How are we framing this — close-up portrait, three-quarter body, wide establishing shot?"
7. **Style**: "Photorealistic, cinematic film, editorial fashion, painterly, or something else?"

#### For scene/environment content — ask in order of impact:

1. **Setting**: "Describe the place — what does it look like, and when is it?"
2. **Mood/atmosphere**: "What feeling should hit the viewer instantly?"
3. **Lighting**: "What's the light source and quality?"
4. **Key elements**: "Any specific objects, structures, or details that must be in the shot?"
5. **Style**: "Photorealistic, stylized, concept art, painterly?"

#### For video content — additional questions:

1. **Motion**: "What's moving — the subject, the camera, or both?"
2. **Duration**: "How long? (Short: 3-5s vs. long: 15-60s changes model choice)"
3. **Audio**: "Do you need sound/music, or silent?"

---

### Step 4: Technical Questions (ask only if not obvious)

These can usually be inferred from context, but ask if unclear:

- **Aspect ratio**: "Standard 1:1 portrait, 16:9 cinematic, 9:16 vertical/social?"
- **Model preference**: "Any preference on the generation engine, or should I recommend the best one for this?"
- **Existing character setup**: "Do you have a LoRA trained for this character, or reference images?"
- **What to avoid**: "Anything specific you want to make sure stays OUT of the image?"

---

### Step 5: Confirm and Synthesize

Before generating the prompt, briefly reflect back the vision:

> "Got it. Here's how I'm reading this: [1-2 sentence summary of the concept]. Let me build that prompt."

Then immediately generate the full output below.

---

## Output Format

Deliver all four components, clearly separated:

---

### 🎯 Positive Prompt

[Hand the completed brief to `comfyui-prompt-engineer` when available; otherwise apply the confirmed model-family cues below]

Key rules:
- **FLUX / Kontext**: Natural language, 50-100 words, no quality tags, describe the scene not the face if using identity method
- **SDXL**: Quality tags first, trigger word second, 50-150 words, weighted syntax supported
- **SD 1.5**: Short and tag-based, 30-80 words
- **Wan / Video**: Concise, motion-focused, 20-50 words
- If a LoRA trigger word applies, put it first
- If using an inventory-verified identity method: follow that implementation's facial-description guidance

---

From calebmills99/VUORSE-VORTEX, MIT.

More from calebmills99/VUORSE-VORTEX