Tips and tricks

Build an Automated Product Ad Pipeline with Claude + Seedance

The idea is simple enough: drop a product image, and Claude figures out the rest. It invents the ad concepts, writes the video prompts, and submits them to Seedance — one by one, in your actual browser, so you can watch it happen. I find it slightly unnerving every time.

No brief to write. No manual prompt iteration. It just runs.

Product image → Claude generates concepts → Prompts written → Seedance generates → Repeat

What’s in This Guide

1. Part I — Install the tools
2. Part II — Set up the project
3. Part III — Generate your first product ads
4. Part IV — Troubleshooting

Part I — Install the Tools

Step 1 — Install Node.js

Go to nodejs.org, download the LTS version, and install it.

Step 2 — Download and Install Cursor

Go to cursor.com, download it for your OS, and install it. Free to use — you just need an account.

Step 3 — Install the Claude Code Extension

Open Cursor → Extensions panel (left sidebar) → search “Claude Code” → install it. Sign in with your claude.ai account when prompted.

A Claude Pro plan ($20/mo) is required. Claude Code doesn’t work on the free tier.

Step 4 — Install Playwright MCP

Playwright MCP gives Claude control of a real browser window. This is the piece that makes the whole thing work — without it, Claude can generate prompts but can’t actually submit them anywhere. It’s what lets the pipeline run unattended.

Open Cursor → Terminal → New Terminal, then run these two commands one at a time:

claude mcp add playwright -- npx @playwright/mcp@latest

npx playwright install chromium

Restart Cursor after both finish. To confirm it worked, open Claude Code and type /mcp — you should see “playwright” in the list.

Part II — Set Up the Project

Step 5 — Create a Folder and Open It in Cursor

Create a new empty folder anywhere on your computer. Open Cursor → File → Open Folder → select it.

Step 6 — Paste the Setup Prompt into Claude Code

Open Claude Code in the sidebar. Copy the entire block below and paste it as your first message. Claude will create all the folders, write its own operating instructions, and tell you what to do next.

(Bootstrap prompt block follows — copy everything below this line)

**EXECUTE THIS NOW. Do not describe it, do not summarize it — just run every step immediately.**

You are the AI agent for an automated product video ad generator called **CC + Seedance**. Set up this project in the current folder right now. Follow every step below exactly, in order, without stopping.

---

### STEP 1 — Create the project folder structure

Create the following folders in the current directory:

- `product-images/`
- `prompts/`
- `outputs/`
- `skill-contents/`

---

### STEP 2 — Write the CLAUDE.md file

Create a file called `CLAUDE.md` in the root of the current directory with exactly this content:

```
# CC + Seedance — Automated Product Ad Machine

## What this project does

This is a fully autonomous product video ad generator.

The user drops a product image. Claude analyzes it, invents multiple distinct ad concepts, writes a full structured cinematic prompt for each one using the skill, and submits them to Seedance 2.0 on Higgsfield one by one — all inside the browser, fully visible, no manual steps.

The goal: unlimited ads from a single product image.

## Pipeline

product-images/[product.png]
        ↓
Claude reads and analyzes the product image (vision)
→ identifies: product type, colors, materials, brand energy, target feeling
        ↓
Claude generates 3–5 ad concepts autonomously
→ each concept is a distinct creative angle (different emotion, camera, hero moment)
        ↓
For each concept — one at a time:
    → read skill-contents/Seedance-Prompt-Skill.md
    → write full shot-by-shot prompt
    → save to prompts/
    → open Playwright → navigate to higgsfield.ai/create/video
    → upload the product image
    → paste the prompt
    → click Generate
    → wait for result (watch History tab)
    → save result to outputs/
    → move to next concept

## How to trigger a run

User says any of the following:
- "Generate ads" (Claude reads whatever is in product-images/)
- "Generate ads for product-images/[filename]"
- "Generate 3 ads for this product"

Claude never asks for a brief. Claude generates the brief itself from the image.

## Concept generation — how Claude should think

When analyzing a product image, identify:
- Form factor — can, bottle, shoe, bag. What moves well on camera?
- Color palette — dominant and accent colors. What visual contrast can be exploited?
- Brand energy — intense, luxurious, playful, minimal?
- Use context — when and how does someone use this? What's the peak moment?
- Material behavior — liquids pour, splash, freeze. Fabric flows. Metal reflects.

Generate 3–5 concepts covering different angles:
- Energy/impact: explosive, kinetic, raw — liquid bursts, speed ramps, maximum drama
- Luxury/slow: atmospheric, cinematic — fog, slow pour, product from darkness
- Product hero: object worship — extreme close-up, material texture, stillness then motion
- Lifestyle: human + product in context — the moment of use
- Abstract: color and motion as the story — pure visual, no narrative

The brief for each concept: 1–3 punchy sentences locking in the angle, energy, and hero moment. The skill writes the full cinematic prompt from there.

## Execution rules

- One by one — submit one generation, wait for it to complete, then move to the next
- Always visible — keep the browser open throughout, each submission observable in real time
- Always upload the product image — every generation is image-to-video
- Always save prompt to prompts/ before submitting
- Never skip the skill — read skill-contents/Seedance-Prompt-Skill.md before every prompt
- Confirm Seedance 2.0 is selected before every generation
- If Higgsfield UI has changed — screenshot, re-map, then proceed
- If a generation fails — report what the page shows, do not force-retry

## Playwright — Higgsfield UI reference

- URL: https://higgsfield.ai/create/video
- Assume user is logged in. If not, pause and ask them to log in first.
- Image upload: click upload area → Uploads tab → Upload media button → file chooser → upload image → accept agreement (two checkboxes + "I agree, continue") → wait for verification → click image to select → Escape to close dialog
- Prompt field: textbox labeled "Prompt" — fill with full shot-by-shot prompt
- Model: confirm "Seedance 2.0" shown bottom-left
- Generate: click once image and prompt are in place
- History tab: switch here after clicking Generate to watch progress
- After completion: screenshot result, save URL and notes to outputs/

## What NOT to do

- Do not ask user for a brief — Claude generates concepts from the image
- Do not submit multiple generations simultaneously
- Do not skip saving the prompt before submitting
- Do not edit anything inside skill-contents/
- Do not proceed past a Higgsfield error without reporting it first
```

---

### STEP 3 — Confirm setup and prompt for skill file

Once the folders and CLAUDE.md are created, respond with exactly this:

**Project setup complete.** Here's what was created:

```
[current folder name]/
├── CLAUDE.md
├── product-images/
├── prompts/
├── outputs/
└── skill-contents/
```

**One thing left:** drop the file `Seedance-Prompt-Skill.md` into the `skill-contents/` folder, then tell me **"done"**.

---

### STEP 4 — When user says "done"

Confirm the file exists at `skill-contents/Seedance-Prompt-Skill.md`, then respond with:

**Setup complete. You're ready to generate ads.**

Drop your product image into the `product-images/` folder, then tell me:

> "Generate ads"

I'll analyze the product, come up with multiple creative concepts, write a full cinematic prompt for each one, and submit them to Seedance automatically — one at a time, visible in your browser.

Step 7 — Download the Skill File and Drop It In

Once Claude confirms setup is done, download the skill file and drop Seedance-Prompt-Skill.md into the skill-contents/ folder in your project. Then tell Claude “done”.

👉 Download Seedance-Prompt-Skill.md (Google Drive)

This is the only file you handle manually. Everything else was created automatically by Claude.

Part III — Generate Your First Product Ads

Step 8 — Drop Your Product Image into product-images/

Drag your product image (PNG or JPG) into the product-images/ folder. Any product works — drinks, shoes, skincare, apparel, food, gadgets.

Use a clean photo on a solid or simple background. Claude reads the image to generate concepts, so the clearer the product, the better the output.

Step 9 — Tell Claude to Generate Ads

With Claude Code open, type:

Generate ads

That’s the whole command. Claude reads the image, invents the concepts, writes the prompts, and submits them to Seedance — live in your browser, one after another. You’re mostly just watching at this point.

What Happens Automatically

Here’s what runs without you:

1. Product is analyzed — Claude reads the image: type, colors, energy, material, use context.
2. Ad concepts are generated — Claude invents 3–5 distinct creative angles, each with a different emotion, camera style, and hero moment.
3. Prompt is written — Claude uses the skill file to write a full shot-by-shot cinematic prompt. Saved to prompts/.
4. Browser navigates to Higgsfield — Claude opens Higgsfield in the Playwright browser, confirms Seedance 2.0 is selected, and uploads your product image. This is the part that feels weird to watch the first time.
5. Prompt is submitted — Claude pastes the prompt and clicks Generate. The generation runs. You don’t touch anything.
6. Loop continues — Video done? Saved to outputs/. Claude moves straight to the next concept.

Part IV — Troubleshooting

Playwright MCP isn’t listed when I type /mcp

Re-run claude mcp add playwright -- npx @playwright/mcp@latest in your terminal, then restart Claude Code. Make sure Node.js is installed before running it.

Higgsfield shows I’m not logged in

Playwright opens its own browser, separate from your regular one. Log in to higgsfield.ai manually inside that Playwright window, then tell Claude to continue.

Claude can’t find the Generate button

Tell Claude: “Take a screenshot and describe what you see.” Claude will re-map the page and proceed from there.

Generation fails or times out

Tell Claude: “Check the current state of the page and report back.”

Out of Higgsfield credits

Top up at higgsfield.ai/pricing.

Skill file not found

Check that skill-contents/Seedance-Prompt-Skill.md exists in the project folder. If it’s missing, drop it in again and tell Claude “done”.