Tithi Mondal

Article9 min readUpdated

Screenshot to Code: What It Can and Cannot Do for You

Screenshot to code turns a picture of a screen into front end code. See what goes in, what comes out, what it needs and where the output stops.

By Tithi Mondal, UI/UX Designer

Read inहिन्दीবাংলাमराठीதமிழ்తెలుగుગુજરાતી

Screenshot to Code: What It Can and Cannot Do for You: cover illustration

Suppose you run a coaching centre in Asansol. You have a screenshot of a fee payment page that you like. You want a similar page on your own website by Friday. A developer friend says: "Give the screenshot to an AI tool. The tool will write the code."

That method is called screenshot to code. Screenshot to code means an AI model looks at a picture of an interface and writes the front end code that draws a similar screen. Front end code is the part of a website or app that people see and touch.

The best known open-source example is screenshot-to-code, a project published on GitHub under the account abi. I read the project's README to write this article. I am not reviewing the tool from use. I am explaining what the tool is, what the tool is for, and where I think the tool fits in a design process.

Checked in November 2026. Tools change quickly, so confirm details on the project's own page.

What goes in and what comes out?

A picture of an interface goes in, and front end code for that screen comes out. According to the project's documentation, the tool converts screenshots, mockups, Figma designs and screen recordings into code using AI.

What you giveWhat you get back
A screenshot of an existing pageCode that draws a similar page
A mockup or a Figma designCode in the stack that you choose
A screen recording of a website in useA functional prototype, according to the README

A stack is the set of coding tools that a page is written with. The README lists these stacks:

  • HTML with Tailwind
  • HTML with CSS
  • React with Tailwind
  • Vue with Tailwind
  • Bootstrap
  • Ionic with Tailwind

HTML is the language that describes what is on a page. CSS is the language that describes how the page looks. Tailwind and Bootstrap are CSS frameworks, which means ready-made styling systems. React and Vue are tools that developers use to build larger apps. You do not need to choose the stack alone. Ask your developer which stack the rest of your product uses, and choose that one.

So the phrase "image to HTML" is a fair description of the simplest use. You give one image. You receive the code for one page.

What does the tool need before it runs?

The tool needs a key for at least one paid AI model provider, and a developer who can set the tool up. A key is a secret code that lets the tool use an AI model on your account. The README says you need at least one model provider key, from OpenAI, Anthropic or Google Gemini.

The README adds these points:

  • A Gemini key is required for the video mode.
  • A key for a service named Replicate is used for image generation and image editing.
  • The project has a back end part and a front end part. A developer starts both.
  • The project can also be started with Docker. Docker is a tool that packages software so that the software runs the same way on any computer.
  • The README says open-source models run through Ollama are not recommended, because of poor-quality results.

The README also points to a hosted version on a separate website, and says that the hosted version is run independently by a company. If you want to avoid the setup, check the current plans on that website. The README states that the code is released under the MIT licence.

The work happens in a few simple steps:

  1. You give the tool an image of a screen.
  2. You choose the stack.
  3. The tool sends the image, with instructions, to the AI model.
  4. The model writes the code, and the tool shows the result.
  5. You or your developer review the code and correct the code.

How could a founder or small team use it this week?

Use the tool to make a first draft of one simple screen, and then judge that draft with a person who knows your users. Do not plan to build the whole product this way in one weekend.

Here is a realistic plan for a team of 2 or 3 people:

  1. Pick one screen that is simple and public, such as a landing page or a pricing section.
  2. Make a clean mockup of that screen first. I would expect a clearer picture to leave the model with less to guess.
  3. Ask your developer to run the tool with the stack that your product already uses.
  4. Open the result on a mid-range Android phone, not only on a laptop.
  5. Write down every difference between the picture and the result.
  6. Decide what to keep. I would expect the layout to be worth keeping more often than the details.

Suppose a tiffin service in Durgapur wants a one-page website with a menu, a weekly plan and a WhatsApp order button. A draft made from a mockup could save the developer the first hours of typing layout code. The owner still has to decide the order of the sections, the wording of the weekly plan, and what happens after a customer taps the button.

One caution about the picture you use. A screenshot of another company's page is that company's design. Use such screenshots to study structure. Do not use such screenshots to copy a brand. Your own mockup is the honest input.

Where does the output fall short of a real product?

The output is one moment of one screen, turned into code. A real product is many screens, in many conditions, for many kinds of people. A screenshot cannot show most of those conditions, so the model cannot know them. These are limits I would expect from the method itself, not faults of this project.

States

A state is a condition that a screen can be in: loading, empty, error, success or disabled. A screenshot shows only one state. Suppose the screenshot shows a fee payment form that is filled in correctly. The code will draw that form. The code will not know what to show when the UPI payment fails, when the OTP expires, or when the network drops. Someone must design each of those states. I have written about them in empty states, loading and error screens.

Responsiveness

Responsive design means the layout adapts to the width of the screen. A desktop screenshot says nothing about how the page should rearrange on a small phone. The model has to guess. A table with 4 columns can become unreadable on a narrow screen. My guide on what responsive web design is lists the decisions that someone has to make.

Accessibility

Accessibility means that people with weak eyesight, people who use a screen reader, and people who use only a keyboard can still use the screen. A picture carries no information about heading order, labels for form fields, focus order or text descriptions for images. Generated code can look right and still fail these checks. A person has to test contrast, labels and tap target size. My guide to accessibility in app and web design explains what to test.

Design system

A design system is the shared set of colours, type sizes, spacing and components that every screen uses. The tool works from the image that you give. If you convert 10 screenshots separately, I would expect 10 slightly different buttons. A developer then has to merge those buttons into one component. Without that step, every later change must be made in 10 places. My guide to a design system for startups explains the minimum set.

Real content

The screenshot shows a name that fits on one line. Real names can be long. Real addresses can need 4 lines. A page in Bengali or Hindi often needs more line height than the same page in English. The generated code knows only the text in the picture.

What are the risks to check before you rely on it?

The main risks are the running cost, the privacy of what you upload, and output that looks finished but is not finished. None of these risks is a reason to avoid the tool. Each risk is a reason to check.

  • Cost of running models. Each conversion uses a paid AI model through your key. Many attempts on many screens add up. Ask your developer to watch the usage page of the model provider.
  • Privacy of your data. The image is sent to an outside AI provider. Do not upload screenshots that show customer names, phone numbers, bank details or KYC documents. Use dummy data in the mockup.
  • Made-up details. A language model can invent text, links or icons that were not in the image, and can misread small text. Read every word of the result.
  • Different results each time. The same image can produce different code on a second attempt. Treat each result as a draft.
  • Code quality. Code that looks correct on screen can still be hard to maintain. A developer must review the code before the code goes into your product.
  • A draft is not evidence. A page that looks complete does not prove that customers understand the page.

What do a designer and real users still have to do?

A designer still decides what the screen should be, and real users still show whether the screen works. The tool starts after those decisions and stops before that proof.

Before the screenshot exists, someone has to choose the flow: which screen comes first, what the screen asks, and what can wait. On a seller app I designed for a marketplace, the order of the verification steps mattered more than the look of any single screen. No screenshot contains that decision.

After the code exists, someone has to design the missing states, set the rules for small screens, check accessibility and bring every screen back to one design system. Then a few real users should try the main task on their own phones. My guide to usability testing on a small budget shows a low-cost way to do that.

My position is simple. Screenshot to code is one way to convert design to code with AI. The method is useful when the design is already decided and the team wants a first draft of the code quickly. The method does not decide the design. AI front end generation saves typing. AI front end generation does not save thinking.

If you want the screens, the states and the handoff file prepared before any code is generated, my UI/UX design service covers that work.

Sources

Frequently asked questions

Is screenshot-to-code free to use?

The code of the open-source project is public on GitHub, and the README states an MIT licence. Running the tool still uses a paid AI model through your own provider key, and the hosted version has its own plans, so check the current details on the project's page.

Can screenshot to code build a full mobile app?

No. The method produces front end code for the screen in the picture. Login, payments, databases, error states and the links between screens still have to be designed and built by people.

Is it safe to upload screenshots of my app to an AI tool?

The image is sent to an outside AI model provider, so remove customer names, phone numbers, bank details and documents first. Use dummy data in any mockup that you upload.

Do I still need a designer if AI can turn an image into HTML?

Yes, for the decisions the image does not contain: the flow, the missing states, small-screen layout, accessibility and one consistent design system. The tool converts a decided design into a draft of code.

Working on something like this?

Tell Tithi what you are planning. You get a clear scope, a fixed price and a timeline before any work begins. I reply to every enquiry within 24 hours.

Your details go only to Tithi and are used only to reply to you. No newsletters, no sharing.

Have a product in mind? Let us design it properly.

Tell me what you are building. I will reply with questions, a rough timeline and a quote.