Cole Medin walks through eleven coding-agent frustrations in a 25-minute tutorial. His recurring method: describe the observed failure, propose a fix, and review where the change belongs in your local skills, rules or hooks.
Verify what the agent actually built
- Premature “done” (1:49): ask what could still fail; use a second reviewer with fresh context for the project’s real failure modes.
- Claims without proof (4:50): request commands and output, screenshots or a working user journey instead of a test-count sentence.
- Passing tests, broken app (6:32): test the interface as a user would; put the journey and evidence in the plan before implementation.
Give instructions an enforceable home
- Ignored hard rule (8:26): move mechanically checkable prohibitions to executable hooks at the action boundary. Keep conventions in rules and workflows in skills.
- Bloated global rules (12:37): remove obsolete guidance and test whether moving a rule to a hook or skill improves the result.
- Context rot (14:01): split work into small sessions; pass a handoff file listing what is done, stuck and tried.
- Lost human control (15:43): make the plan list decisions the agent owns and decisions that need approval.
- Unwanted refactor (17:29): define a focused change and inspect the planned scope before editing.
Put limits around execution
- Usage spikes (19:08): specify when, how many and why to launch sub-agents; assign bounded jobs.
- Overtesting (21:04): set validation depth by blast radius. Item 3 separately sets what user behavior needs testing.
- Tangents (22:43): park an unrelated issue for another task; do not ask the agent to hunt for extra work.
- Caveat: the “90%” payoff is Medin’s heuristic, not an evaluated success rate. The Arcade segment is a disclosed sponsorship; its access claims are separate from the eleven fixes.
“Determine how much testing it does based on the complexity and blast radius. And you determine what exactly it is testing.” — Cole Medin, 22:28