# Browser Image Attachment — Skill

## What
Attach reference images (character sheets, previous panels, style references) to browser-based AI image prompts (Gemini, ChatGPT) for visual consistency across panels.

## Why
Without reference images, every panel is generated blind. Character faces drift after 3-4 panels. Environments change color palette. Objects disappear or transform. Attaching the previous panel + character sheet gives the AI a visual anchor — it sees what it needs to maintain.

## When to Use
- Generating panel N when panels 1 through N-1 exist
- Any panel with a recurring character (face within 5ft of camera)
- Scene continuity (same environment across 2+ panels)
- Style consistency (maintaining the same art direction)

---

## What to Attach

### For Character Consistency
```
ALWAYS attach:
1. Character reference sheet (from objects/ folder)
   - 2x3 turnaround + expression sheet
   - Neutral grey background, no environmental contamination
2. Most recent panel showing that character (from panels/ folder)

OPTIONAL (for complex scenes):
3. Any panel that established the environment this scene is in
```

### For Scene Continuity
```
ALWAYS attach:
1. The immediately previous panel (panel N-1)
2. The first panel that established this location

OPTIONAL:
3. Any reference image for the environment/setting
```

### For Style Consistency
```
First panel of the project: attach style reference images
Subsequent panels: attach panel 1 as style anchor
```

---

## How — Gemini

### Step 1: Prepare Files
Before opening the browser, identify what to attach:
```
Read scene-plan.md for panel N:
  - Which characters appear? → find their sheets in objects/
  - Is this the same location as a previous panel? → find that panel
  - What's the previous panel? → always attach it
```

### Step 2: Upload via File Attachment (Gemini)
Gemini's prompt area has a file attachment button (paperclip or + icon).

```
1. Click the attachment button (find it with read_page or find tool)
   find(query: "upload file button" OR "attach image" OR "add image", tabId: <tabId>)

2. Use file_upload tool to attach the image
   file_upload(paths: ["C:/path/to/character-sheet.png"], ref: <file_input_ref>, tabId: <tabId>)

3. Wait for upload confirmation (thumbnail appears in prompt area)
   computer(action: "wait", duration: 3, tabId: <tabId>)
   computer(action: "screenshot", tabId: <tabId>)

4. Repeat for each reference image (up to 14 in Gemini Nano Banana Pro)
```

### Step 3: Type the Prompt WITH Reference Instructions
After images are attached, type the prompt that tells Gemini HOW to use the references:

```
"Image A is the character reference sheet for [CHARACTER NAME].
Image B is the previous panel showing the same scene.
Generate the next panel maintaining the same character appearance,
environment colors, and art style.

[ACTUAL SCENE DESCRIPTION FROM scene-plan.md]

16:9 widescreen cinematic frame."
```

### Step 4: Submit and Continue Normal Flow
Press Enter → wait → verify → download → cleanup (as per gemini-image-generation.md)

---

## How — ChatGPT

### Step 1: Same File Preparation
Identical to Gemini — identify character sheets + previous panels.

### Step 2: Upload via ChatGPT's Attachment
```
1. Find the attachment/paperclip button
   find(query: "attach file" OR "upload" OR "paperclip", tabId: <tabId>)

2. Use file_upload tool
   file_upload(paths: ["C:/path/to/reference.png"], ref: <file_input_ref>, tabId: <tabId>)

3. Wait for thumbnail to appear
```

### Step 3: Prompt with References
Same pattern — tell the AI which image is what and how to use them.

---

## Attachment Strategy by Panel Number

| Panel | What to Attach | Why |
|-------|---------------|-----|
| 1 | Style reference only | Establishing the look |
| 2 | Panel 1 + character sheets for any characters | Style anchor + character anchor |
| 3-5 | Panel N-1 + character sheets | Continuity chain |
| 6+ (same scene) | Panel N-1 + first panel of this scene + character sheets | Scene + character consistency |
| 6+ (new scene) | Panel N-1 (for style only) + character sheets + new environment ref | Style continuity into new setting |
| After scene change | DO NOT attach old environment panels | Prevents contamination |

---

## File Paths Convention

```
Character sheets:  projects/{id}/objects/{character-name}-sheet.png
Previous panels:   projects/{id}/panels/{style}/panel-{NN}.png
Style references:  projects/{id}/objects/style-reference.png
Environment refs:  projects/{id}/objects/{environment-name}.png
```

---

## Prompt Template with References

```
REFERENCE IMAGES:
- Image 1: Character sheet for {NAME} — maintain this face, hair, build, clothing
- Image 2: Previous panel (panel {N-1}) — maintain this environment, lighting, color palette
- Image 3: [optional] Environment reference — this is the setting

GENERATE:
{Scene description from scene-plan.md}

MAINTAIN:
- Same character proportions and features as Image 1
- Same color palette and lighting as Image 2
- Same art style throughout
- 16:9 widescreen cinematic frame
```

---

## Common Mistakes

| Mistake | Fix |
|---------|-----|
| Attaching too many references (5+) | Stick to 2-3 max per prompt. More = confused output. |
| Using panel from different scene as environment ref | Only attach panels from the SAME location |
| Forgetting to tell the AI which image is what | Always label: "Image A is..., Image B is..." |
| Attaching character sheet with background | Character sheets must be neutral grey — see reference-sheet-system.md |
| Not attaching previous panel | ALWAYS attach N-1 unless it's panel 1 |
| Attaching ALL previous panels | Just N-1 (and first panel of scene if relevant). Not the whole history. |
| Using `form_input` for file upload | Use `file_upload` tool with the file input ref |

---

## Integration with Other Skills

- **frame-chaining.md** — defines "same scene" anchoring logic
- **reference-sheet-system.md** — how to create the sheets you're attaching
- **object-system.md** — reference PNG management
- **nano-banana-prompting.md** — how to write the prompt that goes WITH the attachments
- **gemini-image-generation.md** — the full Gemini workflow (this skill adds Step 2.5)
- **multi-session-coordination.md** — when parallel tabs need different attachments