schliff

Evaluating Claude Code instruction files involves deterministic scoring across seven dimensions: structure, triggers, quality, edges, efficiency, composability, and clarity. This detects degradation and provides actionable recommendations for reliable improvement.

Deterministic scoring for prompt files provides a canonical score by analyzing structural quality across multiple dimensions. It includes anti-gaming safeguards and optional runtime validation when an eval suite is available, ensuring consistent and reliable assessments.

To track prompt degradation over multiple sessions, utilize cross-session memory features that log iteration histories and store prior results. This supports cross-session learning by maintaining a memory of previous scores and measuring progress over time.

Yes, you can score SKILL.md files that reference external scripts and assets. The evaluation supports multi-file skill handling, enabling complex skill trees and handoffs between components like scripts and references during the scoring process.

Generating actionable improvements for Claude Code prompts involves identifying structural gaps through deterministic scoring and proposing atomic changes. The evaluation logs iteration histories for auditability, guiding reliable and measurable prompt enhancement.

You do not need an eval suite to perform structural scoring on instruction files. The seven-dimensional structural scoring works independently, while the eval suite is an optional integration for baseline scoring, runtime validation, and cross-session progress measurement.