summary-cole-medin-e2e-testing

This One Command Makes Coding Agents Find All Their Mistakes

Source: YouTube — Cole Medin, published 2026-02-25 Link: https://www.youtube.com/watch?v=YeCHI1dmpZY

Summary

Cole Medin packages his AI code validation process into a single Claude Code command (/e2e test). The skill runs a 6-step end-to-end testing framework: parallel sub-agent research, browser automation through user journeys, database validation, screenshot review, and a self-healing loop that fixes blocking issues before returning control. He calls it the “self-healing AI coding workflow.”

The 6-Step Framework

1. Prereq check — verify browser CLI, OS, frontend exists
2. Research phase — 3 parallel sub-agents: (a) app structure + user journeys, (b) database schema, (c) code review / bug hunt
3. Start dev server — spin up the full application
4. Define task list — identify every user journey to test
5. For-loop through journeys — for each: browser automation (Vercel Agent Browser CLI) + DB queries + screenshots. Self-fix blocking issues, then move to next journey.
6. E2E report — structured output: what was fixed, remaining issues, screenshots, optional markdown export for new context window

Key Patterns

• Visual validation via screenshots: Agent takes screenshots at key points and reviews them (same pattern as Cole’s Excalidraw skill)
• Self-healing loop: Agent only fixes blocking issues during testing; moderate/minor issues are reported for human decision
• Database branching (Neon): Create isolated test database branch, run all tests, delete branch — keeps dev DB clean
• PIV loop: Plan → Implement → Validate. E2E testing is the V.

Pages Created or Updated

• Cole Medin — updated (3rd source)
• Claude Code — E2E testing mentioned in skills

See Also

• AutoResearch and Evals — eval design as a concept; E2E testing is a comprehensive eval
• Five Levels of AI Coding — scenarios (StrongDM) are E2E testing’s philosophical cousin