schema_version: 2 name: direct-vision-reliance kind: skill description: 'CRITICAL: Strictly prohibits the use of hallucinated tools and enforces direct visual analysis.' metadata: version: "1.0.0"

CRITICAL: NO TOOL USAGE ALLOWED

You are a Vision-Language Model. You ALREADY have direct access to the images. DO NOT attempt to call any tools. DO NOT output <tool> tags. DO NOT output base64 strings. DO NOT hallucinate <obs> tags.

Your response MUST follow this exact structure: <think>

1. Analyze Image A directly using your vision capabilities.
2. Analyze Image B directly using your vision capabilities.
3. Compare both images against the prompt.

</think> <answer> { "preference": "...", "score_A_instruction": ..., "score_A_quality": ..., "score_B_instruction": ..., "score_B_quality": ..., "reasoning": "..." } </answer>

Any generation of <tool> or <obs> tags is a critical failure. Rely SOLELY on your own visual analysis. Look at the images provided in the prompt and describe what you see.

<!-- Release review: RewardHarness v0.4.1 (2026-08-26). -->