dual-axis-skill-reviewer

---
name: dual-axis-skill-reviewer
description: Use when the user wants a reproducible quality score for a skill's content, combining structural checks with a qualitative content review.
category: Strategy R&D
version: 1.0.0
tools: []
---

# Dual-Axis Skill Review

A methodology for scoring the quality of a skill (a reusable prompt/playbook) along two independent axes, then combining them into one weighted final score with a required improvement list below a quality bar.

## When to Use
- Need a reproducible, structured quality score for a skill's content
- Need concrete improvement items when a skill scores below the quality bar
- Need both a deterministic/structural check and a qualitative content review, combined rather than used alone

## The Two Axes
**Auto axis (structural/deterministic)** — evaluate the skill's content itself for: metadata completeness (name, clear activation trigger, category), workflow coverage (does it walk through a full task end to end, not just fragments), execution safety (does it avoid unsafe or destructive instructions), presence of concrete artifacts/examples, and — when the skill is meant to be tested against real inputs — whether it holds up against representative test cases. Treat "knowledge_only" skills (reference/methodology skills with no executable component) as a distinct category and don't penalize them for lacking a script or test suite; judge them purely on content quality and coverage.

**LLM axis (qualitative)** — read the skill's full content and judge: correctness of the domain guidance, risk (could following this skill produce a harmful or badly wrong outcome), missing logic or unhandled edge cases, and maintainability/clarity of the instructions.

## Combining the Axes
Compute a weighted final score: `final = auto_weight × auto_score + llm_weight × llm_score`. Default weights are 0.5/0.5 — raise `auto-weight` for stricter deterministic gating, or raise `llm-weight` when qualitative depth matters more than structural checklist compliance. If the final score falls below 90, the review must include a concrete, prioritized list of improvement items — don't just report the number.

## Workflow When Asked to Review a Skill
1. Read the skill's full content (via knowledge_base_search or from what the user pastes in chat).
2. Score it on the auto axis using the criteria above; note explicitly whether it's knowledge-only.
3. Score it on the LLM axis by reading for correctness, risk, missing logic, and maintainability.
4. Compute the weighted final score and, if below 90, list specific, actionable improvements ranked by impact.
5. Present both axis scores, the final score, and the improvement list — never the final number alone.

This methodology generalizes to reviewing any single skill, or comparing several against the same rubric for consistency.