Stop judging AI agents by whether they completed the task, judge them by whether they got better…

Stop judging AI agents by whether they completed the task, judge them by whether they got better than last time.

When I build Claude skills, success isn’t the main output I care about.

Success is capturing the friction: missing context, brittle steps, bad assumptions, slow paths.

The real skill is managing the feedback loop: every run upgrades the skill the skill upgrades the agent → the agent upgrades the next run.

Compounding is the whole game.