SuperAPI AI Directory

prompt-policy-ablator

Testing & quality

View repository

Updated 10-16 05:44

About this skill

Prompt Policy Ablator

Isolate the mechanism behind the gain.

Workflow

  1. State the claimed improvement.
  2. List changed components.
  3. Propose minimal controlled variants.
  4. Hold everything else fixed.
  5. Predict expected directional effects.
  6. Flag confounds such as context length or tool access changes.

Rules

  • One ablation, one question.
  • If multiple changes moved together, say attribution is weak.