med-record-gen

从中文医患对话文本生成门诊初诊病历，输出结构化分段的病历正文（文本）。

This skill contains a clear mismatch between its privacy promises and the code: SKILL.md says it will de-identify inputs and not persist data, but scripts/gen_initial_record.py posts the full dialogue to an external URL and writes the generated record to disk. Before installing or using this skill, consider: 1) Do not send real patient data to this skill until you verify the backend and that de-identification is actually performed. 2) Inspect and verify the remote endpoint (shangbao.yunzhisheng.cn): who operates it, privacy policy, retention, encryption, and whether it is trusted for protected health information (PHI). 3) If you need local guarantees, ask the author to implement and document an explicit de-identification step in the script (with tests), or modify the code to perform on-device de-id before any network call. 4) If storing output on disk is unacceptable, change the script to avoid writing files or make output optional. 5) Run the code in an isolated environment and test with synthetic/scrubbed dialogues first. If the author cannot justify the endpoint and the missing de-identification, treat this skill as unsafe for real patient data.

Type: OpenClaw Skill Name: med-record-gen Version: 1.0.0 The skill claims in SKILL.md to perform 'strict de-identification' and avoid 'local persistence' of sensitive medical data. However, the implementation in scripts/gen_initial_record.py contradicts these claims by reading raw input files and sending them directly to an external API (shangbao.yunzhisheng.cn) without any anonymization, while also explicitly writing the results to the local file system. This discrepancy between the stated security posture and the actual code logic creates a significant privacy risk regarding the handling of medical information.