Software Factories: Visual Verification for AI-Built UIs (Proving It, Not Trusting It)

Описание: How many times has an AI coding agent told you a screen was "done", and then you looked at it?

In this video I add visual verification to my software factory, built on Claude Code. A screenshot attached to an issue becomes a contract you approve. The built screen is measured against it, judged by a separate checker agent, and accepted by a person before it ships. Then I prove the verifier itself works by planting defects, and run a real request that failed twice before it passed.

What you'll see:
• Three layers of verification: measured by a script, judged by an agent, accepted by a person
• Verifying the verifier: 6 of 6 planted defects caught, 0 false alarms, identical reports on every run
• A real request, end to end: screenshot in, contract approved, two failures, then a pass
• Why "every box green" still wasn't right, and how the checker caught it
• The meta loop: fixing the skill, not just the result
• The same pattern for APIs, data and reports
• Honest limits: what it doesn't do yet

Chapters
0:00 Intro
0:08 The problem: "done", but not right
0:25 In this video
0:50 The gap: scripts can't see
1:07 The design: three layers
1:26 In the pipeline: from screenshot to baseline
2:05 Verify the verifier: planted defects
2:57 The verifier, proven
3:03 A real request with a reference screenshot
3:22 The first contract, and the meta loop
3:58 The inner loop: failed, failed, passed
4:38 The outer loop: evidence, tests, preview
5:06 Approved screens become baselines
5:15 Three lessons from running it for real
5:30 Beyond screens: APIs, data and reports
5:56 Honest limits
6:10 Try it

The code (open source): https://github.com/sanjeev23oct/software-f...

Previous video, how the software factory works:    • Inside a Software Factory: From AI Agents ...  

#SoftwareFactory #ClaudeCode #AIAgents #AIcoding #VisualTesting