Can Any Model Use Skills? Adding Skills to Context-Bench Fresh

Published: November 7, 2025 | Original Post

---

Measuring how well models discover and load specialized knowledge from a skills library.

Key Findings

• Skills improve performance -- Approximately 14.1% improvement in task completion with relevant skills
• Model-agnostic capability -- GPT-5, Gemini, and GLM-4.6 achieve similar gains as Claude models
• Selection challenge -- Performance drops ~6.5% when models must identify the correct skill from a library
• Capability threshold -- Weaker models show negligible improvements, suggesting sufficient reasoning ability is necessary

Skills Concept

Skills are mountable packages of specialized knowledge structured as directories containing metadata files and supplementary resources. Rather than loading all skills simultaneously, agents dynamically mount only relevant skills during task execution.

The process:

1. Select available skills based on metadata
2. Mount relevant ones
3. Execute tasks
4. Unmount when finished

Implications

Skill acquisition works in practice today, enabling agents to extend capabilities online without constant retraining -- crucial for long-horizon, real-world deployments requiring domain adaptation.