NEW Qwen Agent Skill.md (outperforms Anthropic): Trace2Skill

2026-03-29 · Source: Discover AI · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems · Depth: Advanced, long

Summary

Alibaba Q and Application Business Group, ETH Zurich, University of Zurich, Peking University, and Zhejiang University introduced Trace2Skill, a framework for automatically generating agent skill MD files. This system addresses the scalability bottleneck of manual skill file creation and the fragility of existing automated methods. Trace2Skill employs 128 parallel sub-agents to analyze a wide range of trajectory local lessons, distilling common patterns into a single, comprehensive agent skill. The framework significantly improves performance over strong baselines, even outperforming official Anthropic XLSX skills for spreadsheet automation. The study also investigates skill transferability, finding that skills authored by a 122B parameter model generally improve the performance of a 35B model, and a 35B authoring for a 35B model also shows improvement. However, parametric knowledge alone, without environmental interaction, does not yield useful skill content, confirming the necessity of real-world feedback for effective skill generation.

Key takeaway

For Machine Learning Engineers optimizing LLM performance, Trace2Skill offers a robust method to generate highly effective skill MD files. You should consider implementing this parallel analysis approach to overcome the limitations of manual or purely parametric skill creation, especially when targeting complex tasks like spreadsheet automation. This can significantly boost model performance and transferability, particularly when a larger LLM authors skills for smaller models, maximizing existing LLM intelligence utilization.

Key insights

Trace2Skill automates skill file generation for LLMs by parallel analysis of execution traces, outperforming manual and parametric methods.

Principles

• Parallel analysis improves skill distillation.
• Environmental interaction is crucial for useful skills.
• Larger LLMs can author skills for smaller LLMs.

Method

Trace2Skill uses 128 parallel sub-agents to analyze successful and failed execution traces, proposing skill patches. These patches are then merged via inductive reasoning with programmatic conflict prevention into a consolidated skill update.

In practice

• Use Trace2Skill for automated skill generation.
• Prioritize real-world interaction for skill development.
• Consider 35B models as a lower bound for complex tasks.

Topics

• Trace2Skill
• Agent Skill Generation
• LLM Skill Transferability
• Parallel Agent Analysis
• Spreadsheet Automation

Best for: AI Scientist, Machine Learning Engineer, NLP Engineer

Related on AIssential

• Trace2Skill: Distill Trajectory-Local Lessons into Transferable Agent Skills — Trace2Skill distills broad agent experience into transferable, declarative skills via parallel analysis and inductive consolidation.
• A Framework for Evaluating Agentic Skills at Scale — Agent skills significantly enhance LLM behavior and performance, especially for instruction following and making cheaper models competitive.
• Skills on the Fly: Test-Time Adaptive Skill Synthesis for LLM Agents — Dynamically synthesizing task-specific skills from retrieved trajectories improves LLM agent performance without parameter updates.
• SkillBrew: Multi-Objective Curation of Skill Banks for LLM Agents — Principled, multi-objective curation of LLM agent skill banks is crucial for efficiency and self-improvement, moving beyond append-only growth.
• vercel-labs / skills — The "skills" CLI centralizes management of reusable agent instruction sets across diverse coding agents.
• 4 Layer "AI Harness" For LLMs (+54%). Really? — The "AI Harness" improves LLM agent performance by addressing interface mismatches and syntax issues, not by modifying the frozen model.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Discover AI.