How to play: Some comments in this thread were written by AI. Read through and click flag as AI on any comment you think is fake. When you're done, hit reveal at the bottom to see your score.got it
I've always found screenshots on PRs incredibly helpful as a reviewer. Historically I've had mixed success getting my team to consistently add screenshots to PRs, so this tool would be helpful even for human code.
At work, we've integrated claude code with gitlab issues/merge requests, and we get it to screenshot anything it's done. We could use the same workflow to screenshot (or in this case, host a proofshot bundle of) _any_ open PR. You would just get the agent to check out any PR, get proofshot to play around with it, then add that as a comment. So not automated code reviews, which are tiresome, but more like a helpful comment with more context.
Going to try out proofshot this week, if it works like it does on the landing page it looks great.
This would be _extremely_ valuable for desktop dev when you don't have a DOM, no "accessibility" layer to interrogate. Think e.g. a drawing application. You want to test that after the user starts the "draw circle" command and clicks two points, there is actually a circle on the screen. No matter how many abstractions you make over your domain model, rendering you can't actually test that "the user sees a circle". You can verify your drawing contains a circle object. You can verify your renderer was told to draw a circle. But fifty things can go wrong before the user actually agrees he saw a circle (the color was set to transparent, the layer was hidden, the transform was incorrect, the renderer didn't swap buffers, ...).
I had claude build a backdoor command port in the Godot application I'm working on. Using commands, Claude can interact with the screen, dump the node tree, and take screen shots. It works pretty well. Claude will definitely iterate over layout issues.
This is a good point. For anything without a DOM, screenshot diffing is basically your only option. Mozilla did this for Gecko layout regression testing 20+ years ago and it was remarkably effective. The interesting part now is that you can feed those screenshots to a vision model and get semantic analysis instead of just pixel diffing.
playwright can do all of that too. I'm confused why this is necessary.
If coding agents are given the Playwright access they can do it better actually because using Chrome Developer Tools Protocol they can interact with the browser and experiment with things without having to wait for all of this to complete before making moves. For instance I've seen Claude Code captures console messages from a running Chrome instance and uses that to debug things...
Yeah I’ve never seen it capture video before, but if you specify in your `AGENTS.md` that you want to test certain types of workflows, it will take progressive screenshots using a sleep interval or by interacting with the DOM.
Everyone is comparing this to Playwright but it's solving a different problem. Playwright checks structural properties, like does element X exist, is it visible, etc. That's useful but it can't tell you whether the page actually looks right.
I built something similar that takes a screenshot and uses a multi-modal LLM to evaluate it against a design mock. It catches a completely different class of error. The DOM can be structurally perfect and still look nothing like what was intended. Colors wrong, layout shifted, spacing off, components overlapping. No amount of DOM assertions will catch that.
These are two different kinds of gates: structural which are fast and deterministic, and stochastic which are slow but catch things that are completely different. There is very little overlap between the issues, and you want to catch both.
That way I can invest a lot of time getting the mock just right, then let the agents "make it so".
It's the whole tool that's important, not so much how you get screenshots. That's what I'm saying: this is headed in the right direction, it just falls a little short of what I do, where I get tons of value over and above just playwright (or whatever gets the screenshot).
The critical part is that viewed at a high level, this method tests something different, which means it catches different errors.
I'm going the opposite of everyone else is saying.
This is sick OP based on what's in the document, it looks really useful when you need to quickly fix something and need to validate the changes to make sure nothing has changed in the UI/workflow except what you have asked.
Also looks useful for PR's, have a before and after changed.
Useful until your screenshot dims slightly due to display settings or a font renders 1px different on CI. Flaky visual diffs will make you hate this faster than you'd expect.
Exactly. We need more tools like this. With the right model, picking apart images and videos isn't that hard. Adding vision to your testing removes a lot of guess work from ai coding when it comes to fixing layout bugs.
A few days ago I had a interaction with codex that roughly went as follows, "this chat window is scrolling off screen, fix", "I've fixed it", "No you didn't", "You are totally right, I'm fixing it now", "still broken", "please use a headless browser to look at the thing and then fix it", "....", "I see the problem now, I'm implementing a fix and verifying the fix with the browser", etc. This took a few tries and it eventually nailed it. And added the e2e test of course.
I usually prompt codex with screenshots for layout issues as well. One of the nice things of their desktop app relative to the cli is that pasting screenshots works.
A lot of our QA practices are still rooted in us checking stuff manually. We need to get ourselves out of the loop as much as possible. Tools like this make that easier.
I think I recall Mozilla pioneering regression testing of their layout engine using screenshots about a quarter century ago. They had a lot of stuff landing in their browser that could trigger all sorts of weird regressions. If screenshots changed without good reason, that was a bug. Very simple mechanism and very effective. We can do better these days.
I usually ask Claude Code to setup a software stack that can build/run whatever I am working on. Then I let it browse a website or navigate through screens. I also use Playwright to get screenshots of the website I am building. For e.g. apps or whatever application you are building, there should be a way to get screenshots too I guess.
Added benefit is that when Claude navigates and finds a bug, it will either add them to a list for human review or fix it automatically.
Pretty much a loop where building and debugging work together;-)
I'm currently experimenting with running a web app "headless" in Node.JS by implementing some of the DOM JS functions myself. Then write mocks for keyboard input, etc. Then have the code agent run the headless client which also starts the tests. In my experience the coding agents are very bad at detecting UX issues, they can however write the tests for me if I explain what's wrong. So I'm the eye's and it's my taste, the agent writes the tests and the code.
Has anyone actually benchmarked this? I'd expect context window pollution to matter a lot for long sessions, but for short agentic loops it might be noise. Curious if "efficient" means tokens, latency, or just fewer tool calls.
I've been using playwright-cli (not mcp) for this same purpose. It lacks the video feature, I guess. But at least is local and without external dependencies on even more third parties (in your case, vercel). Perhaps you could allow to use a local solution as an alternative as well?
I use AI agents to build UI features daily. The thing that kept annoying me: the agent writes code but never sees what it actually looks like in the browser. It can’t tell if the layout is broken or if the console is throwing errors.
I give agent either a simple browser or Playwright access to proper browsers to do this. It works quite well, to the point where I can ask Claude to debug GLSL shaders running in WebGL with it.
Agreed. Anthropic added a plugin accessible under `/plugins` to CC to make it even easier to add MCP Playwright to your project. It automatically handles taking screenshots.
It's not perfect though - I've personally found CC's VL to be worse than others such as Gemini but its nice to have it completely self contained.
This project desperately needs a "What does this do differently?" section because automated LLM browser screenshot diffing has been a thing for a while now.
> often the playwright skill will verify using DOM API instead of wasting tokens on screenshots
So... Bypassing the whole "sees what it actually looks like in the browser. It can’t tell if the layout is broken" parent commentator is talking about? Seems worse, not better.
All the power to you if you build a product out of this, I don't wanna be that guy that says that dropbox is dead because you can just setup ftp. But with Codex/Claude Code, I was able to achieve this very result just from prompting.
I'd love to see an agent doing work, then launching app on iOS sim or Android emu to visually "use" the app to inspect whether things work as expected or not.
Something like OpenAIs agent mode where it drives a mouse and keyboard but against an emulator should be doable. That agent mode is BTW super useful for doing QA and executing elaborate test plans and reporting issues and UX problems. I've been meaning to do more with that after some impressive report I got with minimal prompting when I tried this a few months ago.
That's very different from scripting together what is effectively a whitebox test against document ids which is what people do with things like playwright. Replacing manual QA like that could be valuable.
Appium's been doing screenshot-based mobile assertions since 2012 or so. The emulator integration works fine — you just need the agent to drive ADB or the iOS sim CLI. Nothing new technically, just wiring.
Yeah, the plan fragmentation is painful. We ended up just running two separate Claude subscriptions for a while — felt dumb but it was cheaper than the combined plan. The Chrome extension is nice but you lose all the context from your coding session.
Yeah Claude/Cursor already have tools to access the browser. What I’m missing is a tool to inspect iOS simulator the same way. Is there a tool for that yet? The Xcode MCP wasn’t really helpful.
I built something like this for native application, so that I could get automated feedback loop for the agent instead of making screenshots manually etc. Problem I found is that AI agent understands nothing of the UI. If you tell it "Make buttons evenly spaced", sure it will space them evenly, but without care for the context they are placed in. You have to describe the image yourself and still you'll find it having hard time understanding what's going on. I very much abandoned the idea of AI driven UI development as it is not there yet. I tried with GPT 5.2. Maybe newer models have improved.
I use Claude Code Sonnet and Opus for this and it works pretty well. Not perfect but good enough. At some point, I will have but to go in and tweak layout but like 99% of the functionality and layout is done for me.
This is really cool. Have you thought of maybe accessing the screen through accessibility APIs? For Android mobile devices I have a skill I created that accesses the screen xml dump as part of feature development and it seems to work much better than screenshots / videos. Is this scalable to other OS's?
This is actually interesting. Feels like we’re moving from “generate UI” to “validate UI,” which is a completely different problem. Curious how you handle edge cases where something looks correct but breaks in interaction?
Not to pile on, but I was using Claude Code through the native application and it started doing exactly this on its own, side by side with my prompt, running the server and taking screenshots in the native app. Claude also just launched its own browser control, and while it will take time to mature, I assume any AI company will have this feature in their crosshairs.
From a product design perspective, this looks pretty cool!
This is basically what antigravity (Google’s Windsurf) ships with. Having more options to add this functionality to Open code / Claude code for local models is really awesome. MIT license too!
I am fed up of getting gaslit by coding assistants. "Your AI agent says it's done." really is a problem! Nice packaging here.
I built something similar[0] a few months ago but haven't maintained it because Codex UI and Cursor have _reasonable_ tooling for this themselves now IMO.
That said there is still a way to go, and space for something with more comprehensive interactivity + comparison.
Great to see this but exe.dev (not sponsored but they are pretty cool and I use them quite often, if they wish to sponsor me that would be awesome haha :-]) actually has this functionality natively built in.
but its great to see some other open source alternatives within this space as well.
Taking screenshots and recording is not quite the same as "seeing". A camera doesn't see things. If the tool can identify issues and improvements to make, by analyzing the screenshot, that's I think useful.
> It’s not a testing framework. The agent doesn’t decide pass/fail. It just gives me the evidence so I don’t have to open the browser myself every time.
From the OP, i don't think this is what is meant for what you are saying.
I read it in the same vein as saying that a sub's sonar enables "seeing" its surroundings. The focus is on having a spatial sensor rather than on the qualia of how that sensation is afterwards processed/felt.
We ran into this exact issue with cursor-based agents rewriting CSS and breaking layouts without realizing it. Adding a screenshot diff step in the loop caught regressions we'd never have caught from DOM checks alone. The visual gap between "technically correct HTML" and "looks right" is real.
