discover-improve

What problem does it solve?

Iteratively improve a blueprint's discover-confidence score by applying deterministic fixes + LLM-driven semantic fixes via a halt-loop (ralph-loop-style autonomous iteration). Mirrors /plan-improve but for blueprints.

Core Features & Use Cases

• Deterministic Phase A fixes to clean prose (e.g., hedge words and loopholes) and mark missing/cited material.
• Phase B LLM-driven refinements to address structural gaps, citation density, and plan alignment.
• Integrated with the /discover-confidence scoring cycle to re-score after each iteration and continue until a target verdict is reached.
• HALT-loop orchestration that enforces cycle rules, safety checks, and progress validation.

Quick Start

Invoke the skill with a blueprint slug to run automated deterministic fixes, then a halt-loop will apply iterative semantic fixes until the target verdict is reached.