Repeated failures waste tokens
Boris as quoted by Theo argues that fixing each occurrence costs tokens and can leave other cases missed. A lint rule, check step, or routine can instead automate the recurring class of busywork. Theo notes that this approach is familiar engineering practice rather than an idea invented for agents.
Make the corrections automatic
When you or an agent corrects the same class of issue repeatedly, turn that correction into an executable check. Choose a lint rule, a check gate, or a small script that catches the recurring pattern. Let the check run on each change, and review it as carefully as the code it checks.
Automate corrections that keep returning in reviews
Use this for recurring style, safety, or architecture corrections that currently return as repeated review comments. The relevant pattern is the same kind of busywork appearing across changes and requiring another individual fix. An isolated correction without a recurring class does not yet supply the reason this tip gives for automation.
Overly broad checks can reject good changes
Boris's efficiency argument is not accompanied by a measured result for replacing repeated corrections with checks. A wrong or overly broad check can silently block good changes, so executable rules need code review. The idea still applies when the check captures the recurring issue without rejecting changes that should be allowed.