Set Up Your Own Static Ad Generator

If you run a small brand and you've ever stared at a blank Canva file at 11pm trying to ship one more ad creative for tomorrow, this guide is for you.

The Static Ad Generator is a framework that takes a brand name, a website URL, and a few product photos, and hands you back twenty production-ready static ad images in about an hour. No design tool. No copy block. No back-and-forth with a freelancer. You give Claude the brand info, approve a couple of permission prompts, and walk away while it works.

By the end of this article, you'll have:

• An Anthropic account with the Claude Code app installed and signed in
• A FAL.ai account with credits loaded and an API key set on your machine
• Python 3 installed and ready
• The Static Ad Framework downloaded, opened, and configured
• Your first brand folder set up with product photos
• A finished gallery of twenty ad images you can review, pick winners from, and ship

This is written for users who are new to AI tools, new to anything that runs in a terminal, and a little nervous about the words "API key." That's fine. We're going to walk through every single step, including the parts that other tutorials skip, like signing up for the right accounts and getting the right credentials before you ever try to run anything.

What this framework actually does

Before we touch any setup, here's the mental model. The framework runs in three phases, all driven by Claude.

Phase 1: Brand Research. You give Claude your brand name and your website URL. Claude searches the web for everything it can find about your brand: fonts, colors, packaging, the look of your current ads on Meta, your founding story, your competitors. It writes all of that into a single document called brand-dna.md. This becomes the playbook for every ad it creates. Takes two to five minutes.

Phase 2: Prompt Generation. The framework ships with twenty proven static ad templates. Things like a headline ad, an offer callout, a stat callout, a before-and-after split, a customer review card, a side-by-side comparison. Claude reads your Brand DNA from Phase 1, then fills in every template with your brand's exact fonts, colors, headlines, and copy. It saves the result to a file called prompts.json. Takes about thirty seconds.

Phase 3: Image Generation. A small Python script reads prompts.json, sends each prompt to GPT-Image (the image model from OpenAI we'll use to actually render the ads), and saves the finished image files to a folder. Each image gets generated one at a time. The script also makes you an HTML gallery so you can scroll through all twenty ads in your browser. Takes thirty to sixty minutes depending on speed.

That's it. Three phases. You answer one prompt to start ("generate ads for [brand], URL is [url]"), and the framework does the rest.

What you need before you touch anything

This is the part most tutorials skip and then beginners get stuck thirty minutes in. We're not going to skip it. There are six things you need lined up before you try to generate a single ad. Knock these out in order and you won't hit a wall later.

2.1 Sign up for Anthropic and install Claude Code

Anthropic is the company that makes Claude. The desktop app that drives this entire framework is called Claude Code, and it's a free download once you have an account.

1. Go to claude.com/claude-code in your browser.
2. Click the option to sign up if you don't have an account yet. You can sign in with Google, with Apple, or with an email address. Use whatever you'll remember.
3. Once you're signed in, download the Claude Code Desktop App for your operating system. There's a Windows installer and a Mac installer.
4. Run the installer the same way you'd install any normal app. On Windows, double-click the .exe file and follow the prompts. On Mac, drag the Claude Code icon into your Applications folder.
5. Open Claude Code. The first time you launch it, it will ask you to sign in. Use the same Anthropic account you just made.

2.2 Sign up for FAL.ai

FAL is a service that hosts AI image models so people can use them through code. We're using FAL because it gives us access to GPT-Image, the image model the framework was built around. You'll need an account and you'll need to load it with a small amount of credits.

1. Go to fal.ai in your browser.
2. Click the sign-up button in the top right. You can sign up with Google or with an email and password. Either works.
3. Verify your email if it asks you to.
4. Once you're signed in, you'll land on the FAL dashboard. Take a quick look around. The two pages we care about are "Keys" and "Billing." Both are reachable from the left sidebar or the top nav, depending on the layout when you visit.

2.3 Get your FAL API key

An API key is a long random string that proves you're you when your computer talks to FAL. Think of it like a password that only your scripts use. Anyone who has this string can spend your FAL credits, so you treat it like a password.

1. From the FAL dashboard, go to fal.ai/dashboard/keys.
2. Click the "Add key" button in the top right of the API Keys page.
3. A small dialog called "New Key" opens. It has two fields:
   • Scope: leave this set to API. That's the default and it's the right one for this framework. The "API" scope lets the key run models and call inference endpoints. It can't manage your billing, other keys, or your account, which is exactly what you want for a key that lives on your computer.
   • Description: type a name you'll recognize later, like static-ad-framework. This is just for your own reference so you know which app the key belongs to if you create more later.
4. Click the purple "Create Key" button at the bottom of the dialog.
5. FAL will now show you the key. It looks like fal_xxxxxxxxxxxxxxxxxxxxxx. Copy the entire string immediately. Paste it into a notes app or a password manager so you don't lose it.

2.4 Top up your FAL credits

FAL is pay-as-you-go. You load credit, and image generation deducts from that balance.

1. Go to fal.ai/dashboard/billing.
2. Add a payment method. Credit card or debit card.
3. Top up with $10 to start. That's enough for roughly 50 to 60 high-quality ads, which is two or three full 20-template runs with room to experiment.

2.5 Install Python 3

Python is a programming language. You don't have to write any. The framework includes a small Python script that does the actual work of sending prompts to GPT-Image and saving the resulting images. The script is already written. You just need Python on your machine so it can run.

On Windows:

The newer Windows installer uses the Python Install Manager. The flow is a little different from older guides you might find online, but it's actually simpler. PATH is handled for you automatically. There's no checkbox to remember.

1. Go to python.org/downloads.
2. Click the big yellow button at the top of the page to download the latest Python 3 installer for Windows.
3. Open the file you just downloaded. A Windows app installer dialog appears titled "Install Python Install Manager?". The publisher is "Python Software Foundation". This is the official installer.
4. Leave "Launch when ready" checked at the bottom-left of the dialog, then click the blue "Install Python" button.
5. Wait for it to finish. The dialog closes automatically when it's done.
6. A small terminal window pops up next, titled "Python installation manager configuration helper". It asks: "Windows is not configured to allow paths longer than 260 characters... Update setting now? [y/N]". Type y and press Enter to enable long path support. You may see a Windows administrator prompt. Approve it. Your computer might prompt you to restart.

On Mac: Python 3 comes preinstalled on modern macOS. You probably don't need to do anything. If you want to confirm, open Terminal (Spotlight search "Terminal") and type python3 --version. If you see a version number, you're set.

On Linux: Python 3 is usually preinstalled. If not, install it with sudo apt install python3 python3-pip on Ubuntu or Debian.

2.6 Get the framework files onto your computer

The Static Ad Framework is a folder of files packaged as a zip archive (static-ad-framework.zip). You can download the zip file here:

Once the download finishes:

1. Find the zip file in your Downloads folder.
2. Right-click it and choose "Extract All" on Windows, or double-click it on Mac.
3. Move the resulting folder somewhere you'll remember. We recommend:
   • On Windows: C:\Users\YourName\static-ad-framework
   • On Mac: ~/static-ad-framework (your home folder)
4. Open the folder. You should see files including START-HERE.md, README.md, a skills folder, and a brands folder. Good. That means your unzip worked.

2.7 Get your product photos ready

This is the part that affects the quality of every ad you generate. The framework sends your real product photos to GPT-Image as a reference, so it can render the actual product (your packaging, your label, your colors) inside every ad. The better your input photos, the better your output ads.

You want one to three photos of your product. Three is the sweet spot.

• Photo 1: front of packaging. Clear, well-lit, label visible, on a clean background.
• Photo 2: angled shot. Same product, three-quarter angle so the model gets a sense of depth and shape.
• Photo 3 (optional): lifestyle shot. The product in context. On a counter, in a hand, beside a coffee. Helps with lifestyle ad templates.

Format: PNG, JPG, or WebP. Don't worry about being a professional photographer. Decent phone photos in good light work fine. Avoid heavily filtered, blurry, or low-resolution images.

Setting up the framework in Claude Code

Everything you needed is now in place. We're going to open the framework folder inside Claude Code, install one Python dependency, set your FAL key permanently, and create your first brand folder. Take your time on this section. Each step takes a minute or two.

3.1 Open the framework folder in Claude Code

1. Launch the Claude Code Desktop App.
2. The first time you open it, or any time you start fresh, it'll ask you to choose a folder to work in. Pick the static-ad-framework folder you unzipped.
3. If Claude Code is already open in another folder, go to File → Open Folder (or your platform's equivalent) and choose static-ad-framework.
4. Claude Code will load. It automatically reads the skills/SKILL.md file inside the framework, which is the framework's instruction manual for Claude. You don't have to do anything to make Claude "learn" the framework. It's automatic.

3.2 Install the Python dependency

The image generation script needs one tiny Python library called requests. Instead of opening a terminal yourself, just ask Claude to do it.

In the Claude Code chat box, type exactly:

Press enter. Claude will explain what it's about to do, then ask for permission to run a shell command. Approve the permission prompt. Claude will run the install, and you'll see output confirming it worked. Wait until Claude says it's done before moving on.

3.3 Set your FAL API key on your computer

Now we tell your computer where to find your FAL key. We're setting it as something called an "environment variable," which is just a name-value pair your computer remembers. The Python script looks for one called FAL_KEY when it runs.

Open the notes app where you saved your fal_… key. Then in the Claude Code chat box, type:

Replace fal_xxxxxxxxxxxxx with your real key. Send the message. Claude will show you the exact command it's about to run (using setx on Windows, or updating ~/.zshrc on Mac). Approve the permission prompt.

3.4 Create your first brand folder

Each brand you make ads for lives in its own folder under brands/. This is where Claude saves your Brand DNA, your generated prompts, and your finished ad images.

For this walkthrough, we're going to use a fictional brand called Lumora Skin, a small skincare company that sells a vitamin C serum. You can replace it with your real brand at any point. The steps are identical.

In the Claude Code chat box, type:

Approve the permission prompt. Claude creates the folders and confirms.

Now drop your product photos into brands/lumora-skin/product-images/ using your file manager (File Explorer on Windows, Finder on Mac). Suggested file names:

• product-front.png
• product-angle.png
• product-lifestyle.png

The exact names don't matter. The script picks up whatever's in that folder.

Generating your first batch of ads

Setup is done. We're going to run the whole pipeline now. One message kicks off all three phases.

In the Claude Code chat box, type:

(Replace with your actual brand name and URL when you do this for real.)

Send. Now sit back. Here's what's going to happen, in order.

Phase 1: Brand Research (2 to 5 minutes)

Claude starts searching the web for everything it can find about Lumora Skin. It looks at the website. It looks for the brand on the Meta Ad Library to see what current ads look like. It searches for the founding story, the design agency that built the brand, the color palette, the fonts, the packaging photos, who the competitors are.

You'll see Claude narrating each step in the chat. It might say things like "Searching for Lumora Skin brand colors" or "Reading the homepage to extract the voice and tone." This is normal. You don't have to do anything except watch.

When Phase 1 finishes, Claude saves a file called brand-dna.md into brands/lumora-skin/. Open it if you're curious. It's a structured document with sections like Brand Overview, Visual System, Photography Direction, Product Details, and Ad Creative Style. This is the playbook every ad will follow.

Phase 2: Prompt Generation (about 30 seconds)

Claude reads the Brand DNA, then walks through the twenty templates one by one and fills them in with your brand's specifics. Things like the brand's hex colors, the headline copy that fits its voice, the right product name, the right offer language. It saves the result to brands/lumora-skin/prompts.json.

If you open prompts.json, you'll see twenty entries, each with the full prompt that's about to be sent to GPT-Image. You can edit anything in this file before Phase 3 if you want to tweak copy or descriptions.

Phase 3: Image Generation (30 to 60 minutes)

Claude copies generate_ads.py from skills/references/ into the brand folder, then runs it. The script does the following:

• Reads prompts.json
• Uploads your product photos to FAL so the model can use them as reference
• For each of the twenty prompts, sends it to GPT-Image, waits for the response, and downloads the finished image into outputs/01-headline/, outputs/02-offer-promotion/, and so on
• Saves the prompt text alongside each image in a file called prompt.txt so you can see what generated what
• When all twenty are done, builds an index.html gallery page

Claude will show you the script's progress live in the chat. If you see a permission prompt for running the script, approve it. If anything fails partway through, the script saves what it already generated, so you don't lose work.

Reviewing your gallery and picking winners

When Phase 3 finishes, Claude will tell you it's done and point you to the gallery file. The gallery lives at brands/lumora-skin/outputs/index.html.

To open it:

• Open File Explorer (Windows) or Finder (Mac)
• Navigate to your static-ad-framework folder, then brands/lumora-skin/outputs/
• Double-click index.html

It opens in your default browser. You'll see all twenty ads laid out in a grid, with the template name labeled under each one. Scroll through. Pick three to five favorites. Discard the rest.

Each individual ad image lives inside its own template folder, like outputs/01-headline/headline_v1.png. To use one, just drag it out of the folder into wherever you upload ads (Meta Ads Manager, your email tool, your ecommerce platform).

Don't be surprised if not every ad is a winner. That's the point of generating twenty in one batch. Some will look great immediately, some will need small tweaks, and some you'll throw out. The math still favors you. A 30 percent hit rate on twenty ads gives you six usable creatives for $5. Good luck getting that from a freelancer.

Cost optimization and quality tiers

The framework runs on GPT-Image's "high" quality tier by default. That's what you want for the ads you actually ship. But during testing, you don't always need top quality. Here's how the math breaks down.

Run type	Approximate cost	When to use
One ad at high quality	$0.15 to $0.25	Final production ads
Full 20-template run, high quality	$3 to $5	Real campaigns
Full 20-template run, medium quality	$1.50 to $2.50	Iterating on new templates or testing copy variations

To run cheaper while testing, ask Claude:

To regenerate just specific templates without rerunning all twenty:

Already-completed images aren't re-billed. The script checks what already exists and skips it.

Troubleshooting common problems

If your first run hits a snag, it's almost certainly one of these. Most fixes are 30 seconds.

"FAL_KEY not set" or auth errors

Your key isn't set, or it's set wrong. To check, ask Claude:

If it comes back blank or doesn't start with fal_, set it again from Section 3.3. On Windows, remember to fully quit and reopen Claude Code after using setx.

"python: command not found" or "pip: command not found"

On Windows, this usually means the Python Install Manager finished but the configuration helper got closed before PATH was set. Re-run the installer from python.org/downloads. When the configuration helper terminal opens, let it finish (type y at the long-path prompt and approve the admin prompt). If you're still stuck, restart your computer so the PATH change takes effect, then close and reopen Claude Code. On Mac and Linux, try python3 and pip3 instead of python and pip.

Claude says it can't find the skill

Your project root in Claude Code is set to a parent folder, not to the framework folder itself. Use File → Open Folder and pick static-ad-framework directly. The skills/ folder must be at the top level of what Claude Code sees.

Phase 3 fails partway through

The script saves images as it goes. Re-run with only the templates that didn't finish. For example, if templates 1 through 12 saved but 13 onward didn't:

Generated ads don't look like your product

Add more product reference photos. The model anchors better with two or three angles than with one. A front shot plus a three-quarter angle plus a lifestyle shot covers most use cases.

Generated text is misspelled or garbled

Make sure you're running at "high" quality, not medium or low. Fine text rendering breaks down at lower tiers. Also, the framework wraps every headline in straight double quotes inside the prompt because that pattern produces the most reliable text output. If you've been editing prompts manually, don't strip the quotes.

FAL says "out of credits"

Top up at fal.ai/dashboard/billing. Add another $10 or more.

Going further: custom templates and reusing brand DNA

Once you've shipped a couple of brands through the framework, two patterns will save you a lot of time.

Add your own ad templates

The twenty templates that ship with the framework are a starting point. They're not sacred. Every time you spot a winning ad on Meta that you'd like to copy the structure of, write a new template and add it to skills/references/template-prompts.md.

Templates use bracketed placeholders like [BRAND_NAME], [PRIMARY_COLOR], [HEADLINE], [PRODUCT_NAME], that Claude fills in during Phase 2 from your Brand DNA. Look at the existing twenty templates to see the pattern, then write your own. Save. Next run uses them.

Your template library is the long-term value here. Anyone can call an image API. A curated collection of ad formats that consistently work is what makes this framework valuable to your business.

Reuse Brand DNA for new products in the same brand

Once you've run Phase 1 for a brand, you have a brand-dna.md file. You don't have to regenerate it for new products in the same brand. Just tell Claude:

Claude reuses the existing Brand DNA, regenerates prompts for the new product, and runs image generation. Cuts your per-launch time roughly in half.

Mining real customer reviews for ad copy that actually converts

Here's the upgrade that takes your ads from good to genuinely high-performing. The auto-generated copy from Phase 2 is functional. It sounds like decent ad copy. But it's also generic. The ads that convert best on Meta are the ones that quote real customers, in their actual words.

The workflow is simple:

1. Pull customer reviews from your product page, your Amazon listing, your Trustpilot page, or wherever your reviews live. A tool called Apify is the easiest way to scrape reviews from most sites without writing any code. Search for the relevant Apify "actor" for the platform you need (Amazon Reviews Scraper, Trustpilot Reviews Scraper, etc.) and follow the actor's instructions.
2. Save the reviews as a .txt or .csv file inside your brand folder.
3. Before running Phase 3, ask Claude:
   Inject real customer languageRead the reviews in customer-reviews.txt and update prompts.json so review-style templates use real customer quotes instead of generic copy. Keep the quotes verbatim.
4. Run Phase 3.

GPT-Image renders quoted text verbatim, which is the whole reason we wrap headline copy in straight double quotes throughout the framework. So when a real customer writes "I literally feel my skin glowing within a week," that exact sentence ends up rendered cleanly inside your ad. That's a different level of authenticity than any AI-written copy.

The repeatable process from here on

Once you've done one brand, every brand after it gets faster. Here's the entire workflow for new brands going forward, in three lines:

1. Create a folder under brands/.
2. Drop one to three product photos into product-images/.
3. Send Claude the message: Generate ads for [brand], URL is [url].

That's the whole loop. Twenty ads for around five dollars and an hour of compute. If you're running multiple brands, or shipping new product launches every month, this scales. The setup work you just did is one-time. The brand work is repeatable.

Alternative setup: using VS Code instead of Claude Code Desktop

If you already use VS Code (Microsoft's code editor) for other work, or you'd rather have an integrated terminal in the same window as your files, this path is for you. The end result is identical to Path A. The only difference is which app you're in.

A.1 Install VS Code and the Claude Code extension

1. Download VS Code from code.visualstudio.com.
2. Install it like a normal app.
3. Open VS Code, click the Extensions icon in the left sidebar (or press Ctrl+Shift+X on Windows or Cmd+Shift+X on Mac), search for "Claude Code" by Anthropic, and click Install.
4. Sign in with your Anthropic account when prompted.

A.2 Open the framework folder

Go to File → Open Folder and pick the static-ad-framework folder you unzipped earlier.

A.3 Open the integrated terminal

Go to View → Terminal, or press the backtick key while holding Ctrl on Windows or Cmd on Mac. A terminal panel opens at the bottom of VS Code. It's already inside the framework folder.

A.4 Install the Python dependency

In the terminal, run:

On Mac or Linux:

On Windows:

A.5 Set your FAL key

On Mac or Linux, for the current terminal session only:

To make it permanent, add that exact line to ~/.zshrc (Mac) or ~/.bashrc (Linux). Restart the terminal so it picks up the change.

On Windows PowerShell, for permanent storage:

Then quit and reopen VS Code so the new variable is picked up.

To verify it worked, run echo $FAL_KEY on Mac or Linux, or echo $env:FAL_KEY on Windows. You should see your key printed.

A.6 Open the Claude Code panel and continue

Click the Claude Code icon in the VS Code sidebar to open the chat panel. From here, the steps are identical to Section 3.4 onward. Create your brand folder, drop in product photos, and ask Claude to generate ads.

If you prefer to run the Python script directly from the terminal instead of asking Claude, you can. Once Phase 1 and Phase 2 have created prompts.json:

Swapping the image model: GPT-Image to Nano Banana Pro

The framework defaults to GPT-Image because it renders quoted text verbatim better than any other model right now, and that's critical for ad copy. But there's an alternative on FAL called Nano Banana Pro (Google's image model) that has different strengths. It's better at photorealism and character consistency. It's weaker at clean text rendering. For lifestyle or editorial templates without much copy, you might prefer it.

To swap the image model, edit skills/references/generate_ads.py:

• Change the endpoint from https://fal.run/openai/gpt-image-2 to https://fal.run/fal-ai/nano-banana.
• Change the edit endpoint from https://fal.run/openai/gpt-image-2/edit to https://fal.run/fal-ai/nano-banana/edit.
• Update the request payload. Nano Banana takes aspect_ratio as a string ("1:1", "4:5") instead of an image_size dictionary. It also doesn't have the low/medium/high quality tier system.

Run a few templates after the swap before kicking off a full batch. The two models render text differently, so you'll want to confirm your headline copy still reads cleanly.

You're done. What now?

You just set up an end-to-end ad creative pipeline. Most agencies charge thousands of dollars to deliver what you can now produce in an hour for the cost of a coffee.

The first run is the hardest. You'll spend most of the time on accounts, installs, and getting your head around the workflow. After that, every brand is just three steps and a coffee break while Claude works.

If you take one thing from this article, take this: the framework's value isn't the twenty templates. It's the loop. Brand research, prompt generation, image generation, gallery, repeat. The templates can be replaced. Your library will grow over time as you spot winning ad formats on Meta and add them. The customer reviews you mine for copy will become more specific to your brand. The Brand DNA gets sharper each time you re-run it.

Run it on your own brand first. Ship the ads. See what converts. Then start adding your own templates. That's the long game.