478 GUI skills push GPT-5's desktop success rate up 15.6 points

Notes on CUA-Skill: Develop Skills for Computer Using Agent (arXiv:2601.21123) — Tianyi Chen, Yinheng Li, Michael Solodko, Sen Wang, Nan Jiang, T. Cui, Junheng Hao, Jongwoo Ko, Sara Abdali, Suzhen Zheng, Leon Xu, Hao Fan, and 3 more · January 2026

Note published July 23, 2026 · written by SkillFed’s research pipeline from the paper above · how these notes are made

AI-assisted notes · reviewed by SkillFed Skill retrieval

CUA-Skill turns human computer-use know-how into 478 atomic, parameterized skills across 17 Windows applications — File Explorer, Excel, Word, Chrome, VS Code, PowerPoint, and more. Each skill pairs a natural-language intent with an argument pool and a parameterized execution graph, a directed graph whose nodes are control states and edges are base GUI actions gated by UI predicates. A separate composition graph encodes which skills can legally follow which, turning a flat skill list into a navigable map of higher-level tasks. Built on top, CUA-Skill Agent handles the runtime work: hybrid lexical-and-semantic skill retrieval (Qwen3-Embedding-0.6B, top-5 candidates) narrows the catalog to candidates, argument instantiation treats enumerable and open-ended inputs differently, and a memory log of past attempts drives failure recovery — falling back to raw primitive actions when nothing cataloged fits.

The library's payoff scales with the backbone model — it isn't a fixed lift. Adding CUA-Skill to Qwen3-VL-32B raised task success from 6.5% to 11.8% (+5.2 points); to GPT-4o from 19.6% to 28.1% (+8.5 points); to GPT-5 from 34.6% to 50.3% (+15.6 points) — same 478 skills, three different lifts. Wired into the full CUA-Skill Agent on GPT-5, the system reaches 50.3% single-attempt and 57.5% best-of-three on WindowsAgentArena, a new state of the art ahead of AgentS3's 56.6% best-of-three and well clear of OpenAI's Operator at 37.4%, while invoking only 117 of the library's 478 skills and using fewer execution steps than the systems it beats. Human performance on the same benchmark sits at 74.5%, so a real gap remains.

Key numbers

Skill library scale	478 skills across 17 Windows apps
GPT-5 success rate, no skills → with skills	34.6% → 50.3% (+15.6 pts)
WindowsAgentArena, best-of-three	57.5% (new SOTA)
Human baseline on WindowsAgentArena	74.5% vs agent's 57.5%
Skills actually invoked in eval	117 of 478

What this means for skill authors

The structural bet is the interesting part: instead of prompting a model to reason out low-level GUI actions from scratch on every task, CUA-Skill pre-compiles that reasoning once and reuses it as a graph. The composition graph is the piece most agent frameworks skip — an explicit map of which skills can legally follow which, rather than leaving sequencing to whatever the model infers from a system prompt.

The ablation table is the number worth remembering if you build or vet skills: the same 478-skill library adds +15.6 points to GPT-5 but only +5.2 to Qwen3-VL-32B. A skill isn't a fixed multiplier on capability — the backbone sets the ceiling, and a library tuned against one model's blind spots may barely move a weaker one. Test any nontrivial skill against more than one target model before trusting its reported lift.

It also surfaces a coverage problem every skill catalog eventually hits: 478 skills exist, but only 117 ever got invoked across the benchmark. A large library isn't automatically a useful one if retrieval only ever surfaces a sliver of it. That's the case for search and vetting that rank skills by task fit — not catalogs that just keep growing.

Skills related to this research

wiki-retrieve wiki-retrieve upgrades vault search from page-level to chunk-level granularity, combining contextual prefixes, BM25 sparse retrieval, and dense reranking to reduce retrieval failures. Setup is opt-in with built-in privacy gates: run fully on-machine with synthetic prefixes, or opt into Anthropic API or claude CLI tiers for better context generation.★ 10,007 azure-openai-2025 Azure OpenAI 2025 equips you with the latest models—GPT-5, GPT-4.1, o3, and o4-mini—plus Sora video generation and image tools. Get deployment commands, Python and .NET SDK recipes, SKU selection strategies, and patterns for on-your-data retrieval and vector search.★ 49 agentsop-multi-tenant-rag This skill encodes security-first isolation patterns for RAG systems serving multiple tenants, customers, or permission scopes. It teaches agents to enforce tenant filtering at the vector store query boundary—the only place where isolation truly matters—and pairs vendor-specific primitives (Pinecone namespaces, Weaviate multiTenancy, Qdrant payload filters, pgvector RLS) with framework adapters like LlamaIndex and LangChain. Activate when designing, reviewing, or debugging any retrieval pipeline where the corpus spans more than one tenant.★ 219 Alibabacloud Bailian Rag Knowledgebase This skill enables retrieval from Alibaba Cloud Bailian Knowledge Base through HTTPS API calls. It automates API key management, handles knowledge base discovery and selection, and supports multi-document search with reranking to find the most relevant results for your queries.★ 198

Related notes

References

CUA-Skill: Develop Skills for Computer Using Agent (arXiv:2601.21123)
WindowsAgentArena benchmark
Agent-S2 (computer-use agent baseline)
UFO-2 (computer-use agent baseline)
OpenAI Operator (computer-use agent baseline)