ads-generate

这段内容是审核时保存的快照。外部来源才是完整且最新的版本。

---
name: ads-generate
description: "AI image generation for paid ad creatives. Reads campaign-brief.md and brand-profile.json to produce platform-sized ad images using banana-claude. Requires banana-claude (v1.4.1+) with nanobanana-mcp configured. Triggers on: generate ads, create images, make ad creatives, generate visuals, create ad images, generate campaign images, make the images, generate from brief."
user-invokable: false
---

# Ads Generate: AI Ad Image Generator

Generates platform-sized ad creative images from your campaign brief and brand
profile. Uses banana-claude as the image generation provider.

## Quick Reference

| Command | What it does |
|---------|-------------|
| `/ads generate` | Generate all images from campaign-brief.md |
| `/ads generate --platform meta` | Generate Meta assets only |
| `/ads generate --prompt "text" --ratio 9:16` | Standalone generation without brief |

## Environment Setup

**Required before running:**

- Requires banana-claude (v1.4.1+) with nanobanana-mcp configured
- Run `/banana setup` to configure API key and MCP
- Fallback: if banana is not available, use `scripts/generate_image.py` (deprecated)

If banana-claude is not installed, this skill will display setup instructions
and stop. It will never fail silently.

If banana-claude is unavailable, alternatives include: OpenAI gpt-image-1 ($0.040/image), Stability SD 3.5 ($0.065), or Replicate FLUX.1 Pro ($0.055). Configure via ADS_IMAGE_PROVIDER env var.

## Process

### Step 1: Verify banana-claude

Verify banana-claude is installed (run `/banana setup` to check). If not installed,
display setup instructions and exit.

### Step 2: Locate Source Files

Check for:
- `campaign-brief.md` → primary source for prompts and dimensions
- `brand-profile.json` → brand color/style injection (optional but recommended)

**If campaign-brief.md is found**: Use `## Image Generation Briefs` section as the
generation job list.

**If no campaign-brief.md**: Enter standalone mode (Step 2b).

#### Step 2b: Standalone Mode

Ask the user:
1. Generation prompt (what should the image show?)
2. Target platform (to set correct dimensions)
3. Output filename (optional)

Then skip to Step 5.

### Step 3: Read Provider Config

Load `~/.claude/skills/ads/references/image-providers.md` to confirm:
- Active provider pricing (show user the cost estimate)
- Rate limits for current tier
- Batch API availability

### Step 4: Read Platform Specs

For each platform in the campaign brief, load the relevant spec reference:
- `~/.claude/skills/ads/references/meta-creative-specs.md` (Feed/Reels/Stories specs + safe zones)
- `~/.claude/skills/ads/references/google-creative-specs.md` (PMax/RSA/YouTube generation-ready specs)
- `~/.claude/skills/ads/references/tiktok-creative-specs.md` (9:16 only + safe zone overlay)

Plus `~/.claude/skills/ads/references/platform-specs-cross.md` for cross-platform aspect-ratio + character normalization.

### Step 5: Prepare banana Configuration

Create banana brand preset from brand-profile.json if one does not already exist
at `~/.banana/presets/{brand-slug}.json`.

Select banana domain mode based on campaign brief content:
- **Product**: e-commerce, packshots
- **Editorial**: brand awareness, lifestyle
- **Cinema**: video thumbnails, dramatic
- **UI/Web**: app install, SaaS
- **Portrait**: testimonials, people

### Step 6: Spawn Visual Designer Agent

Spawn the `visual-designer` agent using the Task tool with `context: fork`,
passing the selected domain mode and preset name.

The agent will:
- Parse the image generation briefs from campaign-brief.md
- Inject brand colors and mood from brand-profile.json
- Use banana-claude with the configured domain mode for each asset
- Save to `./ad-assets/[platform]/[concept]/` directory structure
- Write `generation-manifest.json`

### Step 7: Validate with Format Adapter

After the visual-designer completes, spawn the `format-adapter` agent
with `context: fork` to validate dimensions and report missin