hle-reasoning-wrapper_skill

Overview

This skill wraps Humanity's Last Exam (HLE) questions in a strict Chain-of-Thought (CoT) reasoning scaffold to produce step-by-step answers that match a required format. It enforces a two-part output (Thought and Answer) so downstream models provide explicit reasoning and a final, concise response. Use it to standardize logic, traceability, and format compliance when answering HLE-style prompts.

How this skill works

The skill formats an input question into a structured prompt that instructs the model to produce an explicit Thought section followed by a final Answer. After the model responds, the skill validates the output to ensure both sections exist and that the Thought contains stepwise reasoning leading to the Answer. If validation fails, it flags or rejects the output so you can re-prompt or post-process.

When to use it

• When you must enforce step-by-step Chain-of-Thought for HLE questions.
• When answers must follow a strict 'Thought' then 'Answer' format.
• When traceability of reasoning is required for review or grading.
• When automating validation of model outputs before publishing.

Best practices

• Keep prompts concise and explicit about the two-section output (label 'Thought' and 'Answer').
• Limit the Thought to logical steps that directly support the final Answer.
• Use validation feedback to iteratively refine the prompt if outputs deviate.
• Treat validation failures as signals to re-run the model with clearer constraints.

Example use cases

• Grading HLE exam responses where examiners require stepwise justification.
• Building pipelines that log model reasoning for audit or debugging.
• Training or evaluating models on structured CoT compliance.
• Integrating into educational tools that teach structured problem solving.