The below document is a quick 1 pager I wrote for BoomiAI Weekend Hacks.

AI Tampermonkeying - 11th August 2025

Purpose

This documents details a specific AI product idea to “home-cook”¹ enhancements to existing web applications. I request BoomiAI approval to implement this on 30th August at Boomi AI Weekend Hacks BLR :)

Background

Tampermonkey

Tampermonkey² is a cross-platform browser extension with a large userbase (> 10 million folks) that people use to enhance the functionality of their most frequently used webpage. This is done via writing Javascript scripts called userscripts and running it via the browser extension. Tampermonkey also makes it easy to find and install userscripts created by other users, thus letting you access a library of commonly used customisations and enhancements for your favourite web pages.

The typical user of tampermonkey is “tech-comfortable” but not necessarily a software developer. They use a small number of pre-made scripts for quick convenience. They primarily want userscripts that just work for their use case without needing to meddle or modify installed scripts. These range from common use cases like downloading media from Instagram and Twitter, and UI enhancements for Amazon, to downloading reports from ancient & obscure internal dashboards that make you question your career choices.

The Tampermonkey ecosystem operates on open source principles: power users create userscripts for personal needs and share them openly. This creates predictable gaps. Common use cases get covered, but niche domain-specific needs don’t. The flexible distribution model creates sprawl - scripts come from anywhere - different providers, GitHub gists, random forums - raising security concerns. Many scripts are half-baked and unmaintained, with nuanced bugs that break across browser versions. Despite 10 million users, the ecosystem lacks proper stewardship. Users love the concept but face real friction

AI Generated Software

One of the widely used applications of LLMs has been code generation. Models like Claude Sonnet 3.5 have been useful and cost effective for writing accurate functions from well defined and detailed specs. AI generated code often works on the first try, but when it doesn’t, it usually gets there in the first few tries when enhanced with tools in the LLM workflow (like a python interpreter or playwright automation).

Code generation isn’t solely used by developers anymore. Many “non-technical” users now use tools like Loveable and Replit in an end-to-end manner, making AI coding accessible and letting them generate decent-sized software applications as long as they can clearly describe what needs to be built and handle light debugging. The key insight here is that these users aren’t building complex production software serving thousands of users - they’re building new greenfield applications for more constrained use cases.

Opportunity

The opportunity is giving users much greater ability to customize the webpages they use daily through AI-enhanced Tampermonkey. The key insight is that far more users want to enhance existing webpages in simple ways rather than build entirely new software applications (the target of Replit or Loveable).

Users should be able to implement customization ideas quickly by describing them in natural language. For example: wanting a price history widget on Amazon product pages so you know you’re getting a real deal during Diwali. You describe this requirement in plain English, specify using pricehistory.app as the data source, and get a working userscript that displays the widget automatically. Since users are non-technical, the system validates userscripts in a sandbox or on the users’s webpage before presenting it. This approach also handles fixing broken Tampermonkey scripts, enhancing them further, or security reviewing them without users touching JavaScript code and knowing a lick of Javascript.

The aspiration is transforming casual Tampermonkey users into power users, bringing many more people into the ecosystem, and making webpage customisation more powerful and fun.

Recommendation

Principles

Core design heuristics for the AI userscript system

• Outcome-focused: Users want webpage enhancements, not code. The userscript is the means, not the end.
• Fast iteration: Generate working enhancements within 5 minutes of describing the need.
• Contextual interface: Skip chat. Let users highlight webpage elements and describe changes directly where they want them.
• Safety-first: Always validate generated scripts in sandbox before deployment.

High Level Solution

TODO

---