998: How to Fix Vibe Coding

Wes and Scott kick off the discussion by highlighting the challenges with current AI coding practices, such as unpredictability and the issues caused by unreliable outputs. They emphasize the importance of integrating deterministic tools—like code quality analyzers, linters, and bug trackers—to make AI more reliable, predictable, and maintainable. The conversation underscores the importance of turning subjective 'vibe' coding into fact-based, logical processes to prevent errors and improve client trust.

The hosts explore various tools that analyze code quality and identify issues such as duplication, unused dependencies, and complexity hotspots. They recommend tools like jscpd.dev for detecting duplicate code across multiple languages, and fallow.tools for comprehensive static analysis including dead code detection and complexity measurement. Wes discusses the utility of these tools in helping AI generate better code by focusing on maintainability and clarity, and stresses their speed and configurability, as well as the potential for auto-fixes with caution.

The discussion shifts to methods for finding and using components efficiently through Storybook’s AI integrations, which provide documentation and usage examples, improving AI’s accuracy in component selection. They also delve into bug detection—highlighting Sentry's CLI and Spotlight tools for real-time error tracking, root cause analysis, and performance monitoring both in development and production environments. These tools aid in maintaining high code quality and ensuring bugs are caught early.

Scott and Wes review the critical role of formatting tools such as Vite+ VP Check, ESLint, StyleLint, and custom utilities like Clint, which enforce code consistency and style standards automatically. They discuss strategies to integrate these tools into their workflows—especially in CI/CD pipelines—to catch issues early and ensure cleaner, more maintainable code. The conversation emphasizes how these tools make AI-generated code more predictable and aligned with best practices.

The hosts explore advanced tools for testing UI components and simulating user interactions, such as agent-browser, Chrome DevTools MCP, and Lightpanda. They discuss the importance of headless browsers in QA automation, performance testing, and capturing screenshots. Wes raises concerns about resource consumption and potential misuse, mentioning the relevance of using these tools responsibly in automated workflows and scraping scenarios.

Wes and Scott highlight tools like Dex, Beads, and Context7, which manage tasks, to-dos, and document access within AI workflows. These tools help formalize tasks, enable task dependencies, and improve process predictability—making AI-driven project management more structured and transparent. They discuss how integrating these workflow tools can help enforce the use of quality checks at each step, especially on commits.

The conversation concludes with strategies for ensuring AI adheres to deterministic patterns. Techniques include embedding rules into agents.md files, creating custom ESLint plugins, and setting up scripts that run at commit or deployment. Wes notes that explicit instructions and custom plugins significantly improve reliability, and Scott emphasizes the role of documentation and clear expectations for AI behavior.