/react-native-skia

Eval strategy for this skill

This skill ships with:

• evals/evals.json for output-quality checks
• evals/trigger-train.json for description tuning
• evals/trigger-validation.json for held-out trigger validation
• evals/rubric.md for human or LLM review

How to iterate

1. Run the trigger sets multiple times and compute trigger rates.
2. Improve the description using only failures from the train split.
3. Re-check on the validation split.
4. Run the output-quality evals with and without the skill (or against the previous skill version).
5. Use the rubric to compare visual polish, architecture choice, and performance reasoning, not just syntax.

What this skill should outperform

A strong run should consistently:

• choose retained mode, Picture, Atlas, or shader approaches for the right reasons
• keep Reanimated state on the UI thread
• avoid createAnimatedComponent / useAnimatedProps around Skia nodes
• remember web bootstrap and snapshot caveats
• address reduced motion for decorative or always-on motion
• produce visually directed output rather than generic examples

High-value failure patterns to watch

• returns valid Skia code but with no aesthetic concept
• uses the wrong render mode for the workload
• animates via React state or JS-thread churn
• forgets GestureHandlerRootView or LoadSkiaWeb
• uses generic text nodes when Paragraph is clearly the right API
• answers "performance" questions without talking about render mode or workload shape