How to build an automation with Claude Code · From a phone photo to a video ad.

The most important insight from the session comes even before touching the tool: do not rush to find problems for the AI. Start with a real problem you already have, and solve it. The problem we built the automation around in session 21 is content creation: every business repeatedly needs professional product images and advertising videos, and this is a manual, slow, and expensive process. Instead, we built a system that takes a single regular photo from your phone and returns a ready video ad. The demonstration is visual, but the working method is not. Exactly the same process works on your email inbox, office workflows, and anything you do repeatedly or that connects several tools together. Claude Code is the programmer sitting next to you: you explain to it in Hebrew what you want, and it writes the code for you.

What you will have at the end of the guide

• Understanding the full workflow: photo to studio photo to storyboard to video
• The exact prompt to paste into Claude Code to build the automation
• How to work with Plan Mode to save tokens, money, and mistakes
• How to generate an API key in Google AI Studio and connect it securely through the terminal
• How to connect a video model like Seedance 2 via KIE.ai or fal.ai
• How to request a dashboard that turns the entire process into a single click
• Answers to questions raised in the session: security, tokens, sharing, and design

What the system actually does

This is the big shortcut of the new working method: until today, you had to create each frame separately, and then animate each frame with its own prompt. Today, models are smart enough to take a single image with 9 scenes and written instructions, and produce the entire video from it in one shot.

Step 1 · Open a new project in Claude Code

Step 2 · Plan in Plan Mode before building

The full prompt for the automation · Paste in Plan Mode

I want to build an automation that turns a regular product photo into a video ad. Workflow: 1. I upload a product photo. A regular photo from the phone or the web, it does not have to be professional. 2. Send the image to the Nano Banana 2 model via the Google AI Studio API. Clean up the image and turn it into a professional studio photo with a white background. 3. From the clean image, create a single storyboard image in 9:16 format (vertical, suitable for Instagram) divided into 9 separate, numbered scenes, like an advertisement. Below each scene, write a number and a short description with director instructions: what happens in the scene, camera angle, and movement. 4. Stop and show me the storyboard for approval before continuing. 5. After I approve, send the storyboard to the Seedance 2 tool via fal.ai or KIE.ai (whichever is available), and create an ad video based on the instructions written on the frames. Additional requirements: - I will enter the API keys securely through the terminal, not in the chat. Open a terminal for me and give me the command when we get to it. - If something fails (credits, connection, invalid key), display a clear error message explaining exactly what to do. - We will start with a single test scene before running the entire process. Let's plan this before we start building.

Step 3 · Connect the tools with API keys

An API key is like a blank check: it is connected to your credit card, and anyone holding it can generate items at your expense. Do not paste it in chat, do not share it, and do not upload it to the web. If a key is exposed (for example, accidentally pasted in chat), Claude itself will refuse to work with it. The solution is simple: delete the old key and generate a new one.

How to enter the key securely · The most technical action you will take

• Tell Claude: 'I have the API key. Open the terminal for me so I can enter it securely'
• Claude opens a terminal window and gives you a short command. Copy and paste it into the terminal
• Return to Google AI Studio, copy the key, paste it into the terminal, and press Enter (sometimes you do not see what is pasted, that is fine)
• Write to Claude 'I am done'. It checks that the key is saved and the connection works
• If you have worked with the same tool in another project in the past, you can take a shortcut: ask Claude to scan previous projects and pull the existing key from there

Step 4 · Run, get stuck, fix. This is what it is supposed to look like

The AI Master Club opens the door for you: a live session every week, recordings, and a community. ₪47 per month, cancel anytime.

Upgrading: A dashboard instead of a chat

Dashboard prompt

The automation works and I am satisfied with the process. Now I want a user experience instead of working through the chat. Build a simple dashboard (a small site running locally) that centralizes the entire process: 1. Image drop area: I drag one product photo or several photos from different angles. 2. Product description text field: A short mini-prompt that refines the ad. 3. Video settings: Length (4 / 8 / 10 / 15 seconds), quality (720p / 1080p), sound (with / without). 4. A single button: 'Create ad'. 5. Status area showing progress: product analysis, studio photo creation, storyboard, and video. Includes a timer. 6. At the end: A display of all outputs with a download option. Keep the design clean and simple. All the automation logic we already built remains identical, only the wrapper changes.

Where you can take this from here

All the tools from the session

The platform that writes the code and connects all the tools together

All Google models in one place: Nano Banana for images, Veo for video. This is where you generate the API key

Access to many image and video models (including Seedance 2) through a single account and credits

An alternative to KIE: The same idea, a large model library through a single API

The only free software you need to install before starting

Questions raised in the session

Anthropic intentionally decided that its model would excel in language and code, and not be multi-modal like Gemini or ChatGPT. Instead, it created the MCP protocol, which allows Claude to connect to any other software. As soon as you connect it to a tool that knows how to create an image, it creates the image in that tool and retrieves the output for you.

During the building phase, there are fixes and improvements, and that is part of the process. But once the system works completely, all that is left is to input an image and click a button. The entire chain, from image cleanup to the video, runs by itself.

The code sits in a folder you chose on your computer, locally. As long as you have not uploaded the project to the web (for example to GitHub as an open project) and have not given access to other people, there is no exposure at all. The keys also stay with you in the folder.

The key is saved in the same project where you entered it. In a new project, you can ask Claude to scan previous projects and pull the key from there. When generating a key, you can set its validity: an unlimited key will keep working forever, and a monthly valid key is safer if it leaks, but will require renewal.

Either you open the same session from the list on the side and continue, or you open a new session on the same folder. Claude goes over the files, understands the project status, and continues from where you left off. If the context window fills up (the blue sign at the bottom), simply open a new session on the same folder.

Models do not have built-in memory. The solution is a file named CLAUDE.md that reminds Claude in every task who you are, what you want, and how to work with you. There is such a file at the general level that applies to all projects, and you can add a file at a specific project level with dedicated instructions.

AI that writes code tends to use the same colors and fonts, what is called AI slop. The solution: dedicated design skills. On the skills.sh site, there are skills like UI UX Pro Max that immediately improve the fonts, colors, and user experience of anything Claude builds.

Most of these tools are precisely this process behind the scenes: someone built an automation with Claude Code or Codex, connected it to image and video models, wrapped it in an interface, and sells access. When you build it yourself, you control the process, pay only for actual use, and can tailor it precisely to your business.

Want to build such automations with live guidance? In the AI Master Club, we do this every week: we take a real problem from the business, build a system for it live, and fix together everything that gets stuck along the way. The full recording of session 21 is waiting for club members in the content library.

Frequently asked questions

Why doesn't Claude create the images and video itself?

Anthropic intentionally decided that its model would excel in language and code, and not be multi-modal like Gemini or ChatGPT. Instead, it created the MCP protocol, which allows Claude to connect to any other software. As soon as you connect it to a tool that knows how to create an image, it creates the image in that tool and retrieves the output for you.

Is it really automatic? It looks like you are constantly operating it manually

During the building phase, there are fixes and improvements, and that is part of the process. But once the system works completely, all that is left is to input an image and click a button. The entire chain, from image cleanup to the video, runs by itself.

Where is the code saved? Is it secure?

The code resides in the folder you selected on your computer, locally. As long as you have not uploaded the project to the web, for example to GitHub as an open project, and have not given access to other people, there is no exposure. The keys also remain with you in the folder.

Do I need to enter the API key again in every project?

The key is saved in the same project where you entered it. In a new project, you can ask Claude to scan previous projects and pull the key from there. When generating a key, you can set its validity: an unlimited key will work forever, and a key with a monthly validity is safer if it leaks, but will require renewal.

How do I return to a project after I closed Claude?

Either you open the same session from the list on the side and continue, or you open a new session on the same folder. Claude goes over the files, understands the project status, and continues from where you left off. If the context window fills up, indicated by the blue sign at the bottom, simply open a new session on the same folder.

Why doesn't Claude remember me between projects?

The models do not have built-in memory. The solution is a file named CLAUDE.md that reminds Claude in every task who you are, what you want, and how to work with you. There is such a file at the general level that applies to all projects, and you can add a file at the specific project level with dedicated instructions.

The applications I build look generic. How do I improve the design?

AI that writes code tends towards the same colors and fonts, what is called AI slop. The solution is dedicated design skills. On the skills.sh website, there are skills like UI UX Pro Max that instantly improve the fonts, colors, and user experience of everything Claude builds.

I saw paid tools that do exactly this. What is the difference?

Most of these tools are precisely this process behind the scenes: someone built an automation with Claude Code or Codex, connected it to image and video models, wrapped it in an interface, and sells access. When you build this yourself, you control the process, pay only for actual usage, and can adapt it precisely to your business.