Summary
Cole Medin walks through the 11 most common frustrations with AI coding agents and a simple fix for each, aimed at delivering a large payoff for minimal effort. His framing is that builders are now “engineers of the harness”, responsible for the skills, rules, and hooks, which he calls the AI layer, that set agents up for success in a codebase.
The first and most universal problem is an agent claiming work is done when it is not. Medin recommends playing devil’s advocate by asking the agent what would make the implementation fail and what gaps remain against the original issue. He then formalizes that with a separate reviewer skill handed to a different agent for a fresh, critical pass.
A recurring method runs through every tip: describe the problem to Claude Code or Codex, then work with the agent to propose changes to your own AI layer, since the right fix depends on what you build and how you work. He also explains where each kind of guidance belongs. Constraints and conventions go in rules, must-never or must-always behaviors and enforced processes become hooks, and larger guided workflows become skills. One example moves instructions such as avoiding .env files or using bun instead of npm out of CLAUDE.md and into hooks. Each tip has its own chapter for easy navigation.
📺 Source: Cole Medin · Published November 10, 2026
🏷️ Format: Tutorial Demo
