71% of public healthcare skills carry no safety-boundary statement

Notes on An Empirical Study of Agent Skills for Healthcare: Practice, Gaps, and Governance (arXiv:2605.02709) — Gelei Xu, Ningzhi Tang, Xueyang Li, Toby Jia-Jun Li, Zhi Zheng, Wei Jin, Yiyu Shi · May 2026

Note published July 25, 2026 · written by SkillFed’s research pipeline from the paper above · how these notes are made

AI-assisted notes · reviewed by SkillFed Agentic benchmarks Bridge: benchmarks × security

Researchers built the first systematic census of healthcare-focused agent skills — self-contained instruction packages that an agent loads only once a task matches the skill's stated description, a pattern the spec calls progressive disclosure. Starting from 58,159 public skills on ClawHub (an April 2026 snapshot), the team used an LLM classifier to pull 557 healthcare-related skills, then had a second LLM read each skill's full SKILL.md file to annotate ten dimensions covering function, care-cycle stage, intended user, input modality, autonomy level, and clinical decision impact.

The resulting picture is lopsided. Diagnosis and treatment reasoning — the tasks healthcare-agent papers foreground, with 40.0% of papers tagging diagnosis — show up in only 22.4% of skills; workflow automation, research support, and patient monitoring dominate instead. Specialized clinical inputs are nearly absent: medical images and physiological signals each serve as the primary input in just 9 skills apiece, an order of magnitude below structured forms and text chat. Autonomy clusters at L3 ("delegated execution"), but 13 skills combine high autonomy (L4 or L5) with outputs the authors classify as actively driving a clinical decision. Across the whole corpus, only 163 of 557 skills (29%) state an explicit safety boundary.

Key numbers

Healthcare skills identified from ClawHub	557 of 58,159 screened
Diagnosis-support skills, vs. 40.0% of healthcare-agent papers	22.4% of skills
Skills with an explicit safety-boundary statement	163 of 557 (29%)
Medical-imaging / physiological-signal skills (primary input)	9 skills each (<2%)
High-autonomy skills (L4/L5) that drive clinical decisions	13 skills

What this means for skill authors

General technical risk and clinical risk turn out to be different axes. Most of the corpus scores as Safe or Moderate under a standard agent-risk taxonomy — few skills request destructive, financial, or code-execution capabilities — yet a text-only skill with no unusual tool access can still shape what a patient believes about their symptoms or what a caregiver does next. That's a blind spot for any review process that stops at permissions and data access.

For skill authors, the more actionable gap is the disclaimer, not the model. Seventy-one percent of these skills say nothing about their own scope or limits, so most reusable healthcare procedures carry no signal for when they've been invoked outside the context their author intended. That's a metadata problem before it's a safety problem. A vetted catalog can close it by treating scope and boundary statements as a submission requirement instead of an afterthought — the same instinct behind pairing dual-tool scanning with human vetting rather than trusting a permissions scan alone.

Supply is also concentrated and consumer-skewed: a handful of prolific authors account for a third of the corpus, and patients and consumers are the single largest user group at 37.1%, ahead of clinicians or hospital administrators. Specialties that need imaging or lab-value pipelines are barely represented. That's less a verdict on healthcare skills as an idea than an open lane — for the clinicians and domain specialists who aren't publishing yet.

Skills related to this research

Education Tutor Education Tutor provides one-on-one academic support across subjects and grade levels, breaking down difficult concepts through multiple teaching methods tailored to how you learn best. It creates structured study schedules, generates practice problems, and offers test preparation guidance to build real understanding rather than surface memorization.★ 7 Lesson Study Cycle Designer Lesson Study Cycle Designer structures the Japanese jugyō kenkyū process for teacher teams investigating student learning challenges. It generates a research theme, detailed cycle plan with meeting agendas, a research lesson designed for collaborative observation, and protocols focused on case students. The skill grounds itself in evidence from Stigler & Hiebert, Lewis, Perry & Murata, Dudley, and Takahashi & McDougal on how teachers develop shared professional knowledge through structured cycles of planning, teaching, observing, and analysing.★ 472 creator-content-auditor This skill automates the review process for creator-submitted content, comparing each piece against your established brand guidelines and campaign parameters. It helps marketing teams catch misalignments early, maintain consistent brand voice across influencer partnerships, and streamline approval workflows before content goes live.★ 2,469 biostatistics This skill handles survival analysis and time-to-event data common in clinical research, including Kaplan-Meier estimation, Cox proportional hazards regression, and competing risks modeling. It also covers longitudinal data analysis, diagnostic accuracy assessment, and missing data strategies tailored to biomedical studies.★ 869

Related notes

References

Xu, G., Tang, N., Li, X., Li, T. J., Zheng, Z., Jin, W., & Shi, Y. (2026). An Empirical Study of Agent Skills for Healthcare: Practice, Gaps, and Governance. arXiv:2605.02709.
Xu, G., Li, X., Chen, Y., Duan, Y., Wu, S., Yu, H., Chiu, C., Ni, J., Tang, N., Li, T. J., et al. (2026). A Comprehensive Survey of AI Agents in Healthcare. Journal of Biomedical Informatics, 105045.
Ling, G., Zhong, S., & Huang, R. (2026). Agent Skills: A Data-Driven Analysis of Claude Skills for Extending Large Language Model Functionality. arXiv:2602.08004.
Tang, N., Chen, C., Fang, Z., Xu, G., Dhakal, M., Shi, Y., McMillan, C., Huang, Y., & Li, T. J. (2026). Programming by Chat: A Large-Scale Behavioral Analysis of 11,579 Real-World AI-Assisted IDE Sessions. arXiv:2604.00436.
Wu, Y., & Zhang, Y. (2026). Agent Skills from the Perspective of Procedural Memory: A Survey. Authorea Preprints.