AI Model Evaluation Templates

A very short task makes TTFT easier to observe, suitable for chat, agents and real-time interaction.

A classic logic puzzle that checks whether the model gives executable steps instead of only a conclusion.

A short math problem that exposes mistakes around coupon order, units and step-by-step calculation.

Ask for complete runnable code and check completeness, executability and whether the model avoids extra explanation.

Checks whether the model notices the duplicated-element bug and proposes a reasonable fix.

A high-constraint formatting task that checks whether the model adds explanations or breaks JSON.

Five short mixed knowledge questions for a quick read on coverage and concise answering.

Observe creative style, length control and bullet formatting for content production scenarios.

Ask for pure SVG to test visual structure, code completeness and renderability.

One-shot single-file HTML + Canvas animation task for testing runnable, smooth, layered motion effects.

The recently viral 'ball bouncing inside a rotating hexagon' reasoning task: pure Canvas, no network needed, and whether the ball stays inside is visually obvious — strongly separates models' physics and collision reasoning.

One-shot single-file Three.js/WebGL scene task for checking spatial structure, lighting, camera motion and interaction.

Requires the user to paste source text, useful for testing long-input understanding, extraction and structured output.