Vibe coding has opened up exciting new possibilities for the animal protection movement. Anyone can now build a website or app by describing what they want to an AI tool, which writes the code for them. You don’t need to know a programming language. You act as the manager, the tester, and the subject expert, and the AI does the programming.
The term vibe coding was coined in February 2025 by computer scientist Andrej Karpathy. In late 2025, AI models such as ChatGPT and Claude crossed a threshold. They became so good at coding, and made so few mistakes, that they could produce small pieces of software without any programming help from humans.
Where vibe coding works best
Vibe coding tends to work best when someone with deep expertise in a topic, such as grassroots organizing, agriculture, or food, uses AI to build software that shares that expertise with the movement.
The Plant-Based Nudge Calculator from Better Food Foundation is a strong example. Björn Ólafsson, a researcher there, had spent months collecting studies on how menu changes, such as making plant-based meals the default, shift what diners choose. He had Claude Code write a calculator around that research. Now a dining director or an advocate can enter three numbers about a cafeteria and see roughly how many meals a change could shift to plant-based each year, without reading the studies themselves.
That expertise also let Björn check the AI’s work. He wrote the methodology behind the calculator, so he knew what the numbers should be and could tell when the code got them wrong. If you find yourself explaining the same method, calculation, or set of rules to people again and again, that knowledge may be a good candidate for a vibe-coded tool.
Test it thoroughly before others rely on it
Because you’re not writing the code yourself, you can’t check it by reading it. Testing is how you find out whether it works. It is essential that vibe-coded software goes through significant user testing before other people rely on it, both to make sure it works and to make sure it is secure.
Testing a vibe-coded tool usually means:
- Trying every option and every input, including odd ones such as blank fields, zero, very large numbers, and text where a number should go.
- Checking the results against answers you worked out yourself.
- Opening it on a phone as well as a computer.
- Asking a few people it is meant for to use it, without explaining it first, and noting where they get stuck.
Björn and others tested the Nudge Calculator after Claude Code built it, and its public change history shows the kind of problems testing catches. It includes fixes for a carbon result that showed up blank, an evidence rating drawn out of four dots instead of five, effect sizes that needed updating to match Björn’s written methodology, and the calculator getting cut off on phones.
AI coding tools can also write and run their own tests, so ask them to. That still does not replace people using the tool the way your audience will.
Get a security review before publishing anything built on private data
We discourage posting a vibe-coded site publicly if it relies on proprietary or private data underneath, unless you have had a security review from someone with programming knowledge.
Anything a public site loads in a visitor’s browser can usually be seen by anyone who knows where to look, and AI-written code can very occasionally make mistakes here that you will not spot by using the site normally. A small mistake could expose a supporter list, a database password, or the key your organization uses to pay for an AI service. Asking the AI to check its own code for security problems is a sensible first step, but it does not replace a review by a person who can read code.
The Nudge Calculator is a good fit for public posting because everything in it is already public. Its effect sizes come from published studies, and it runs entirely in the visitor’s browser, with no logins and no stored data. If your tool needs private data, keep it on your own computer until someone with programming knowledge has reviewed it.
Protect your brand before you publish
Anything you publish with your organization’s name on it speaks for you. AI-built pages tend to look AI-built: purple gradients, glass panels, rounded cards with a colored stripe down one side, the same font on every heading. Visitors recognize that look, and it can put them off.
Impeccable’s slop catalog lists these habits with examples and fixes. Ask your AI to check your design against it and remove what it finds. Impeccable can also check a project against a file called DESIGN.md that records your fonts, colors, and corner radii, so the design stays inside your own system.
Keep the design in line with your brand as well. Before the AI starts, give it your logo, colors, and fonts, plus a few pages of your site to match, and tell it to use only those. Before the tool goes public, open it next to your website and check that they look like they belong together.
Getting started
Claude Code or OpenAI Codex is where we would start. You describe what you want in plain English, and they write the code, run it, and check that it works. Claude Code needs the Claude desktop app and a paid Claude plan. Codex comes with every ChatGPT plan, but we recommend a paid one for building anything real. For a quick single page, Claude chat’s Artifacts and Gemini’s Canvas can also build simple interactive pages inside the conversation. ChatGPT’s Sites feature can build and host a small website from a description, though as of October 2026 it is a beta on paid plans and is not available in the UK, Switzerland, or the European Economic Area.
Start with something small that uses only public information, like Björn’s calculator. When it is ready to share, you can host a single page free on GitHub Pages and show it inside a page on your own website, which is how Better Food Foundation shows the calculator.
