ProofShot: Autonomous Verification for AI Coding Agents

ProofShot is an open-source CLI tool that allows AI coding agents to record and verify their web development work through an automated session lifecycle. It generates detailed artifacts like videos, logs, and interactive reports to provide transparent 'proof of work' for AI-driven changes. By using agent-specific skill files, it enables AI assistants to autonomously document their tasks and simplify the PR review process.

Key Points

• ProofShot enables AI agents to provide 'proof of work' by autonomously recording and verifying their browser-based tasks.
• The tool generates high-fidelity artifacts including synchronized video, console/server logs, and interactive HTML reports.
• It is designed to be agent-agnostic, offering specific 'skill files' to integrate with popular AI assistants like Claude and Cursor.
• Advanced features include visual regression testing, automatic dev server detection, and multi-language error pattern matching.
• The project focuses on an open-source, no-vendor-lock-in approach to AI-driven development verification.

Sentiment

The community is notably skeptical. While there is genuine appreciation for the concept of visual verification for AI coding agents, the dominant sentiment is that ProofShot doesn't sufficiently differentiate itself from Playwright and other existing tools. Many commenters feel the problem is already solved by first-party solutions. However, a meaningful minority sees real value in the bundled proof artifact workflow and the agent-agnostic approach, especially for non-web platforms.

In Agreement

• The bundled proof artifact (video, screenshots, logs, interactive viewer) for PR review is genuinely useful and reduces the 'agent says it's done but didn't actually verify' problem
• For desktop and native app development without a DOM, screenshot-based visual verification is essentially the only option and tools like this fill an important gap
• An agent-agnostic CLI tool that works with any terminal-based agent (Claude Code, Codex, etc.) is more flexible than IDE-specific solutions
• The concept of shifting from 'generate UI' to 'validate UI' represents an important evolution in AI-assisted development workflows
• Screenshots on PRs are incredibly helpful for reviewers and automating their capture solves a longstanding adoption problem

Opposed

• Playwright and playwright-cli already provide all these capabilities (screenshots, video capture, browser interaction) without additional tooling
• Chrome DevTools MCP and Claude's --chrome flag already let agents interact with browsers natively, making a separate tool redundant
• IDE-integrated solutions like Antigravity (Google's Windsurf) and VSCode already ship with this functionality built in
• AI agents are still fundamentally poor at understanding UI semantics — they can detect structural issues but not whether a layout actually looks right
• This appears to be a 'Not Invented Here' problem where LLM users rebuild existing tools rather than learning what's already available