SkillNet: A Nebula of 200,000 Skills — How AI Moves from Reinventing the Wheel to Inheriting Wisdom

Forum topic · 小凯 · 2026-03-30

Summary

SkillNet is an open skill infrastructure developed by 40+ researchers from institutions including Zhejiang University, Alibaba, Ant Group, and Tencent, designed to give AI agents a systematic mechanism for accumulating, evaluating, and reusing skills. The platform hosts over 200,000 curated skills, each structured as a self-contained folder built around a SKILL.md file with metadata, step-by-step instructions, optional resources, and dependency declarations. Skills are created from execution trajectories, open-source repositories, semi-structured documents, and natural language prompts, then refined through a multi-stage curation pipeline (deduplication, filtering, classification into 10 functional categories, evaluation, and integration). A five-dimension evaluation framework scores skills on safety, completeness, executability, maintainability, and cost-awareness; automated assessors achieved MAE below 0.03 and near-perfect QWK against blind human review. SkillNet also builds a large-scale Skill Graph capturing similar_to, compose_with, belong_to, and depend_on relations, enabling multi-level ontology reasoning. Experiments in ALFWorld, WebShop, and ScienceWorld showed an average 40% reward improvement and 30% reduction in execution steps across DeepSeek V3, Gemini 2.5 Pro, and o4 Mini backbones, demonstrating model-agnostic benefits.

SkillNet: A Nebula of 200,000 Skills — How AI Moves from 'Reinventing the Wheel' to 'Inheriting Wisdom'

> "We are in the era of experience." > —— Richard S. Sutton, 2022

The Experience Dilemma in the AI Era

The post opens with a story: a young programmer spends three days writing a script to extract sales data and generate weekly report emails, only to learn his mentor could have pointed him to a better version written three years earlier, buried in a GitHub corner. This mirrors the core problem facing today's AI Agents: they lack a systematic mechanism for skill accumulation and transfer.

Even advanced agents rely only on in-context learning — once a conversation ends, learned skills dissipate. As Richard Sutton notes, intelligence increasingly depends on efficiently retrieving and adaptively reusing heuristic knowledge from past experience, not learning from scratch. Humans outperform AI not because of faster computation, but because we crystallize experience into reusable patterns.

From Symbols to Parameters to Skills

The post traces three eras of AI knowledge representation:

Symbolic era (1970s–80s): interpretable if-then expert systems, but brittle and unscalable.
Deep learning era (2010s): knowledge as weight matrices — powerful but opaque, non-modular, and fixed after training.
Agentic era (2020s): agents that act, facing the core challenge of converting fragmented experience into durable, composable capability units.

What Is a Skill?

In SkillNet, a skill is a structured, self-contained capability package — like a professional kitchen recipe containing metadata, precise steps, dependencies, alternatives, common failure modes, and variants. Each skill is a folder centered on a SKILL.md file with:

1. Metadata (name, description, usage conditions) 2. Detailed step-by-step instructions 3. Optional resources (scripts, templates, docs) 4. Dependency declarations

Skills operate via a three-step progressive process:

Discovery: load minimal metadata to identify relevant skills
Activation: read the full SKILL.md when a task matches
Execution: follow instructions, optionally running bundled code

Skills are self-documenting — humans can read, audit, and improve them, in contrast to black-box neural networks.

SkillNet's Three Core Modules

1. Skill Creation

Skills are automatically extracted from:

Execution trajectories (recorded agent actions, errors, and resolutions)
Open-source repositories (READMEs, comments, example scripts)
Semi-structured documents (PDFs, slides, Word docs)
Natural language prompts

A multi-stage curation pipeline then applies deduplication (directory comparison + MD5 hashing), rule- and model-based filtering, classification into 10 functional categories (development, AIGC, research, science, business, testing, productivity, security, lifestyle, other) with fine-grained tags, evaluation, and integration into structured skill packages.

2. Skill Evaluation

Each skill is scored on five dimensions, each rated Good / Average / Poor:

Safety: resistance to dangerous operations and prompt injection
Completeness: coverage of key steps, prerequisites, dependencies, constraints
Executability: runtime verification in sandboxed environments (the hardest dimension)
Maintainability: modularity and composability
Cost-awareness: latency, compute, and API cost overhead

Most dimensions are assessed by LLM-based evaluators (GPT-5o-mini). Validation against 200 randomly sampled skills blind-reviewed by three PhD-level annotators showed MAE below 0.03 and QWK near 1.000 across all dimensions.

3. Skill Analysis

SkillNet builds a large-scale Skill Graph capturing four relation types:

similar_to: functionally equivalent, interchangeable skills
compose_with: skills combinable into workflows
belong_to: category/domain membership
depend_on: execution dependencies

This forms a three-level ontology: a Skill Taxonomy, a Skill Relation Graph, and a Skill Package Library (e.g., data-science-visualization), enabling global reasoning over which skills combine, what they depend on, and what alternatives exist.

Experimental Validation: Where the 40% Comes From

Experiments in three text-based simulated environments:

ALFWorld: household navigation and object manipulation
WebShop: simulated online shopping (browsing, search, comparison, purchase)
ScienceWorld: text-based scientific experimentation

Results with SkillNet-equipped agents:

Average reward improved by 40%
Execution steps reduced by 30%

These gains held across DeepSeek V3, Gemini 2.5 Pro, and o4 Mini — meaning the benefits are model-agnostic, stemming from systematic knowledge reuse rather than any specific backbone.

The Essence of Skills

The paper's philosophical claim: skills are intermediate capability units connecting abstract declarative knowledge (knowing-what) with concrete executable procedures (knowing-how). They bridge principle and practice, and are evolvable, composable, and shareable — like human civilization, where each generation stands on the shoulders of the last.

Future Outlook: From Individual Intelligence to Collective Wisdom

SkillNet envisions infrastructure for collective AI intelligence:

Standardized skill interfaces independent of underlying models
An open ecosystem where anyone contributes skills, like open-source software
Automated skill discovery from human experience
Multi-dimensional quality assurance
A relational skill network enabling complex reasoning

The imagined future: a medical AI drawing on the world's top doctors' accumulated diagnostic skills; a coding AI reusing community skill packages for new frameworks; a research AI referencing lessons from all similar past experiments.

References

1. Liang, Y., Zhong, R., Xu, H., et al. (2026). SkillNet: Create, Evaluate, and Connect AI Skills. arXiv:2603.04448. 2. Sutton, R. S. (2022). The Era of Experience. Blog post. 3. Schick, T., et al. (2023). Toolformer: Language Models Can Teach Themselves to Use Tools. arXiv:2302.04761. 4. Qin, Y., et al. (2023). Tool Learning with Foundation Models. arXiv:2304.08354. 5. Shinn, N., et al. (2023). Reflexion: Self-Reflective Agents. arXiv:2303.11366. 6. Wang, L., et al. (2024). A Survey on Large Language Model based Autonomous Agents. arXiv:2308.11432. 7. Wu, Y., et al. (2024). OS-Copilot: Towards Generalist Computer Agents with Self-Improvement. arXiv:2402.07456. 8. Shao, Z., et al. (2024). DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models. arXiv:2402.03300. 9. Zhang, N., et al. (2024). Extract, Define, Canonicalize: An Ontology-based Knowledge Graph Construction Framework. arXiv:2403.00079. 10. Pan, J. Z., et al. (2023). Large Language Models and Knowledge Graphs: Opportunities and Challenges. arXiv:2308.06374.