How AmElmo Built ProofShot to Let AI Coding Agents Verify UI
Built by jberthom · · Source: Hacker News
What they built
AmElmo (the GitHub handle on the repo, which gives no full name) built ProofShot, an open source CLI under the MIT license. The README states the problem as "AI coding agents build UI features blind." An agent can write a feature, but it can't see whether the page looks right, works, or throws errors. ProofShot has the agent test its work in a real browser and leaves behind a bundle of proof that a human can review.
According to the README, it works with Claude Code, Cursor, Codex, OpenCode, Gemini CLI, Windsurf and GitHub Copilot, plus any agent that can run shell commands. It sits on top of agent-browser, which the README credits to Vercel, and it needs no cloud service. The repo page showed 870 stars when we read it.
How it works
Setup is two commands. npm install -g proofshot installs the CLI and agent-browser with headless Chromium. proofshot install detects which AI tools you have and adds a skill or rules file for each one. After that, the README says you tell the agent "verify this with proofshot" and the skill handles the rest.
A session has three steps:
- Start: proofshot start launches your dev server, opens the browser, starts recording and captures server logs.
- Test: the agent drives the browser with agent-browser commands such as open, fill, click and screenshot.
- Stop: proofshot stop bundles everything into a timestamped folder.
That folder holds a video of the full session, screenshots of key moments, a summary report, a timestamped action log, the server log, the browser console output, and an interactive viewer with a scrub bar and timeline that plays the video alongside the logs.
The README also says ProofShot detects errors in server logs across more than 10 languages, including JavaScript, Python, Ruby, Go and Rust. Two more commands round it out: proofshot pr uploads the artifacts and posts a verification comment on a GitHub pull request, and proofshot diff compares screenshots against a baseline folder for visual regressions.
The author is clear about where it fits. The README recommends Playwright MCP or DevTools MCP for live debugging and DOM inspection, and ProofShot for bundled proof artifacts.
How to build something similar
- Give your coding agent a way to open your app in a real browser, and have it click through the feature it changed before it reports the work as done.
- Save evidence from every run in one folder: a recording or screenshots, the browser console output and your dev server logs.
- Put the instructions in a skill or rules file so one short phrase triggers the whole check.
- Attach the evidence to the pull request and look at it yourself before you merge.
The original post
Want to build something like this?
Join the Shipping Skool community free. Paid members get live bootcamp calls every week.
Join Shipping Skool — FreeMore use cases
- How Andrej Karpathy Built HackerTube with Claude Code and Opus 5.5Developer
- How Miguel Ríos Berríos Built Recall to Search Past Agent SessionsDeveloper
- How Adonias Built a CRM in Claude Code with Opus 5.5 and 6 SubagentsDeveloper
- How Kieran Klaassen Built a Webcam Heart Rate App with Gemini 4 ArgonDeveloper