goals
Installation
SKILL.md
Process Goals in Prompt Optimization
Core Principle
Process goals (controllable intermediate actions) provide dense feedback signals; outcome goals (end-result demands) provide sparse, delayed feedback. This asymmetry explains why behavioral prompting dominates direct output demands.
Mechanism: Dense intermediate supervision → stable gradients → reliable optimization
Failure mode: Sparse outcome signal → high variance → reward hacking / hallucination
Goal Typology
| Type | Effect Size | Prompt Analog | Signal Density | Failure Mode |
|---|---|---|---|---|
| Outcome | d=0.09 | "Give the correct answer" | Sparse | Hallucination, reward hacking |
| Performance | d=0.44 | "Achieve high accuracy" | Proxy | Goodhart's Law misalignment |
| Process | d=1.36 | "Think step-by-step" | Dense | Over-specification (rare) |