GOAL.md gives agents measurable goals

This is less a prompt template than a control system for autonomous coding: define "better" numerically, and the agent can grind toward it without gaming the metric.

• –Best for optimization loops like flaky tests, docs quality, link integrity, and performance budgets
• –The split-score pattern is the standout: measure the instrument separately so the agent cannot optimize the wrong thing
• –Atomic commits and re-measurement make the loop safer than a plain "keep iterating" prompt
• –The catch is setup cost: you still have to build the ruler before the agent can use it