skill-eval: data-model-designer 5.5 -> 7.0/10 (eval report + improved SKILL.md)

[jensen-srp]

skill-eval Report: data-model-designer

Original Score: 5.5/10 (Conditional) -> Improved: 7.0/10 (Recommended)

Blind A/B evaluation of data-model-designer by @datadrivenconstruction using skill-eval, an automated evaluation engine that tested 123 ClawHub skills.

Finding

The current SKILL.md is an 11KB Python class definition (ConstructionDataModel). When loaded, the model tries to use these classes, burning +83% time / +113% tokens without producing better schemas than baseline.

Improvement

Rewrote SKILL.md as a behavioral contract:

• Removed all Python code (the model already knows how to code)
• Added mandatory output sections: Requirements Summary, ER Diagram, SQL DDL, Validation Checklist, Sample Queries
• Added SQL style rules and explicit prohibitions
• Result: overhead dropped, score jumped from 5.5 to 7.0

Files added

File	Description
EVAL-REPORT.md	Original blind A/B evaluation
EVAL-IMPROVED.md	Before/after comparison card
SKILL-improved.md	Proposed improved SKILL.md
IMPROVEMENT-LOG.md	Detailed changelog with rationale

Methodology

• Blind A/B: same prompts tested with and without skill loaded
• Two-layer assertions: deterministic checks + rubric-based quality scoring
• Model: Claude Opus 4
• Full methodology

Leaderboard

All 123 evaluated skills: jensen-srp.github.io/skill-eval

---

Filed by skill-eval v0.4.0

[openclaw-barnacle]

Thanks for the pull request! This repository is read-only and is automatically synced from https://clawhub.ai, so we can’t accept changes here. Please make updates on the website instead.