brewui:image-gen

What this skill does

<instructions>

# Image Gen

Generates AI images via Gemini Imagen 4, OpenRouter (Gemini 2.5 Flash Image / GPT-5-image), Z.ai GLM-image, or OpenAI DALL-E 3. Applies anti-AI-slop prompt prefixes per style. Four modes: generate, edit, config, update.

**Arguments:** `$ARGUMENTS`

## Mode Routing

| Mode | Flow |
|------|------|
| generate | Phase 0 -> 1 -> 2 -> 3 -> 4 |
| edit | Phase 0 -> 1 -> 2E -> 3 -> 4 |
| config | Phase 0 -> C |
| update | Phase 0 -> U |

> **CONTEXT-AWARE MODE DETECTION (CRITICAL):**
> Mode is detected from BOTH explicit flags AND natural language context. Priority:
> 1. Explicit flags: `--edit`, `--config`, `--update` -> override everything
> 2. Context analysis of `$ARGUMENTS` text:
>    - Edit signals: "edit this", "modify image", "change the", "add to image" + image path present -> **edit**
>    - Config signals: "setup", "configure", "set key", "add token" -> **config**
>    - Update signals: "check providers", "update models", "latest API" -> **update**
>    - Everything else -> **generate** (this is 99% of cases)
> 3. Default: **generate** — just a prompt, generate the image
>
> **FAST PATH (99% case):** When `$ARGUMENTS` is just a prompt text (no flags, no mode signals):
> - Skip Steps 3-6 in Phase 1 (count, service, style, output questions)
> - Use defaults: count=1, service=gemini, style=photo, output=.claude/reports/images/
> - Go straight to config table (Step 7) + confirmation (Step 8)
> - Only AskUserQuestion if API key is missing
>
> **AGENT INVOCATION:** This skill can be called by agents (not just users). When called from an agent:
> - Treat all provided args as final — do NOT ask for confirmation of values already specified
> - Only AskUserQuestion for truly missing required values (prompt, API key)
> - The config table is still mandatory but confirmation step can be skipped if all params are explicit

> **API KEY PRIORITY** (check in order, first found wins):
> 1. Explicit key in `$ARGUMENTS` text (user pasted inline)
> 2. `.env` in project root (`source .env 2>/dev/null`)
> 3. Shell environment variable
> 4. AskUserQuestion (redirect to Phase C)
>
> **MANDATORY:** Before ANY API call, display the full resolved configuration table. User must see exactly what will be sent.

---

## Phase 0: Parse Arguments

### Step 1: Parse Flags

**EXECUTE** using Bash tool:
```bash
bash "${CLAUDE_SKILL_DIR}/scripts/parse-args.sh" $ARGUMENTS && echo "OK" || echo "FAILED"
```

Output: KEY=VALUE pairs. Store all values.

> **Also scan `$ARGUMENTS` raw text** for inline values not captured by flags:
> - API key pasted in prompt text -> extract and use (overrides `.env`)
> - Service/style mentioned in free text -> treat as flags

| Key | Default | Options |
|-----|---------|---------|
| `PROMPT` | (empty) | Free-text image description |
| `MODE` | generate | generate, edit, config, update |
| `SERVICE` | gemini | gemini, openrouter, openai |
| `STYLE` | photo | photo, illustration, art |
| `COUNT` | 1 | 1-10 |
| `OUTPUT` | .claude/reports/images/ | Directory path |
| `SIZE` | 1024x1024 | WxH format |
| `EDIT_IMAGE` | (empty) | Path to image for edit mode |
| `EDIT_INSTRUCTIONS` | (empty) | Edit instructions text |
| `PROMPT_MISSING` | false | true if no prompt in generate mode |

> **STOP if FAILED** -- check parse-args.sh output for error details.

### Step 2: Route to Mode

| Parsed MODE | Go to |
|-------------|-------|
| generate | Phase 1 |
| edit | Phase 1 |
| config | Phase C |
| update | Phase U |

---

## Phase 1: Validate and Gather

### Step 1: Load Environment and Check API Key

**EXECUTE** using Bash tool:
```bash
[ -f .env ] && set -a && . .env && set +a; bash "${CLAUDE_SKILL_DIR}/scripts/validate-key.sh" "SERVICE_HERE" && echo "OK" || echo "FAILED"
```
Replace `SERVICE_HERE` with the resolved SERVICE value.

> **If FAILED (INVALID):** Redirect to Phase C (config mode). Tell the user: "No valid API key found for {SERVICE}. Let's configure it."

### Step 2: Gather Missing Parameters

If MODE=generate and PROMPT_MISSING=true:

**ASK** using AskUserQuestion:
```
What image do you need? Describe the scene, subject, and mood.
```
Options:
- "Describe your image (e.g., 'a cozy coffee shop at sunset with warm lighting')"
- "Cancel"

Store response as PROMPT.

> **FAST PATH CHECK:** If PROMPT is provided (not missing) AND no explicit --service/--style/--count/--output flags were given:
> → Skip Steps 3-6 entirely. Use defaults (count=1, service=gemini, style=photo, output=.claude/reports/images/).
> → Jump to Step 7 (config table).
> This is the 99% path — user just wants an image from their prompt.

### Step 3: Confirm Image Count (skip on fast path)

**ASK** using AskUserQuestion:
```
How many images to generate?
```
Options:
- "1 (default, fastest)"
- "2-3 (compare variations)"
- "4+ (batch generation, up to 10)"

Update COUNT with the number. Default: 1.

> **Provider limit:** DALL-E 3 supports only 1 image per request. If SERVICE=openai and COUNT>1, generate COUNT sequential requests.

### Step 4: Confirm Service (skip on fast path)

**ASK** using AskUserQuestion:
```
Which image generation service?

| Service | Model | Speed | Quality | Cost |
|---------|-------|-------|---------|------|
| openrouter | Gemini 2.5 Flash Image | Fast | High | ~$0.001/image |
| zai | GLM-image | Fast | **Very High** | ~$0.015/image |
| gemini | Imagen 4 | Fast | Very High | Paid plan required |
| openrouter-gpt5 | GPT-5 Image | Medium | **Highest** | ~$0.01/image |
| openai | DALL-E 3 | Medium | High | $0.04-0.12/image |
```
Options:
- "openrouter (Gemini 2.5 Flash -- cheapest, default)"
- "zai (GLM-image -- flagship Z.ai, high quality)"
- "gemini (Imagen 4 -- high quality, paid plan)"
- "openrouter-gpt5 (GPT-5 Image -- highest quality, ~$0.01/img)"
- "openai (DALL-E 3 -- reliable, most expensive)"
- "Keep current: {SERVICE}"

Update SERVICE if changed. Re-validate key if service changed.

### Step 5: Confirm Style (skip on fast path)

**ASK** using AskUserQuestion:
```
Image style? This controls anti-slop prompt engineering.

- photo: Physically accurate photography -- real lighting, correct anatomy, natural materials
- illustration: Professional illustration -- clean line work, proper color theory, organic imperfections
- art: Consistent artistic medium -- unified brushwork, intentional composition, coherent color temperature
```
Options:
- "photo (realistic photography)"
- "illustration (clean vector/drawn style)"
- "art (painterly/artistic medium)"
- "Keep current: {STYLE}"

### Step 6: Confirm Output Directory (skip on fast path)

**ASK** using AskUserQuestion:
```
Where to save generated images?
```
Options:
- ".claude/reports/images/ (default)"
- "Current directory (.)"
- "Custom path (type your preferred directory)"

Update OUTPUT with chosen path.

### Step 7: Display Resolved Configuration (MANDATORY)

Output this table before proceeding. Do NOT skip this step.

```
=== Image Generation Config ===
| Parameter | Value |
|-----------|-------|
| Prompt | {PROMPT (first 80 chars)}... |
| Service | {SERVICE} ({model name}) |
| Style | {STYLE} |
| Count | {COUNT} |
| Size | {SIZE} |
| Output | {OUTPUT} |
| API Key | {first 8 chars}...{last 4 chars} |
| Est. Cost | {estimate based on service and count} |
================================
```

### Step 8: Final Confirmation

**ASK** using AskUserQuestion:
```
Proceed with generation?
```
Options:
- "Yes, generate"
- "No, change settings"
- "Cancel"

> If "change settings" -> go back to Step 4.
> If "Cancel" -> STOP with message "Image generation cancelled."

---

## Phase 2: Build Payload and Generate

### Step 1: Load Anti-Slop Instructions

Read the anti-slop reference for the resolved STYLE:
Read file: `${CLAUDE_SKILL_DIR}/references/anti-slop.md`

Extract the section matching STYLE (photo, illustration, or art). Store as ANTI_SLOP_PREFIX.

### Step 2: Build Enhanced Prompt

Combine anti-slop prefix with user prompt:
```
ENHANCED_PROMPT = ANTI_SLOP_PREFIX + "\n\n" + PROMPT
```

##