prompt-optimizer

# Prompt optimizer ## Contract | Field | Bound contract | |---|---| | Trigger | User asks to improve, optimize, rewrite, tune, or port a prompt, or to audit prompt text for dated instructions | | Authority | Read-only: no file, VCS, credential, paid, published, deployed, or remote mutation. A proposed diff is chat output; applying it is outside this skill. | | Side effect | Chat output returns either an optimized prompt or an audit report plus a proposed diff | | Done | Optimize mode: shorter prompt validated on holdout cases with one owner per behavior rule. Audit mode: every finding names a pattern and a target-model reason, and the diff carries only high- and medium-confidence hunks. | ## Inputs Required: - The source prompt text or document to optimize or audit - The target task or goal the prompt must serve (optimize mode) Optional: - Mode (optional): optimize (default) or audit. Audit when the request names auditing, dated instructions, cruft, or a model migration; optimize otherwise. When a request asks for both, run audit first and report each result separately. - Known failure cases or error patterns from prior runs - Model family or adapter context (e.g., Claude, GPT-4, Gemini); audit mode reads this as the target model - Evaluation criteria the user already accepts ## Procedure ### Optimize mode (default) 1. **Capture the source.** Record the exact prompt text or document. Note any quoted variable slots, numbered steps, conditional branches, or formatting c...