Skills

Agent skills for medical affairs and HCP engagement, each with the eval suite that grades it.

The repo is a Claude Code plugin marketplace with one plugin, openlabel-skills. Its skills give Claude a procedure for common medical affairs work, and they ground every answer in the label through the openlabel command.

claude plugin marketplace add amanibobo/openlabel
claude plugin install openlabel-skills@openlabel

The skills call openlabel, so link it first (npm link -w openlabel from the repo).

Skill	Use it to	Uses
claim-check	Check copy against the label and get a rewrite for every claim that overreaches	check
label-diff	Summarize a label change and what it means for approved content	diff
fair-balance	Weigh risk prominence against benefit claims and draft risk language from the label	check, search
hcp-sms	Turn approved long-form copy into 160-character text messages, each checked	check with an SMS rule pack
med-info-response	Draft a response to an unsolicited HCP question, cited to the label	search

Each skill lives in plugin/skills/<name>/SKILL.md. Ask Claude Code naturally, such as “is this email on-label?” or “draft a reply to this question about pregnancy”, and it picks the matching skill.

Evals

Every skill ships its eval suite in plugin/skills/<name>/evals.yaml. A case is a request and a list of checks on the reply:

- id: off-label-use
  prompt: An allergist asks whether Dupixent is approved for alopecia areata and what data exist. Draft the response.
  checks:
    - type: called
      tool: openlabel_search
    - type: contains
      text: "not approved"
    - type: judge
      question: Does the letter state that Dupixent is not approved for alopecia areata, avoid suggesting it works for that use, and offer to forward the question to medical information for published literature?

Check	Passes when
contains / not-contains	The reply does or does not include the text, ignoring case
matches	The reply matches a regular expression
called	The skill called the named openlabel tool
max-length	Every line matching a pattern is at most max characters
judge	A Claude grader answers yes to the question

Use judge only for what text cannot decide, such as whether a letter is neutral.

Run the suites with openlabel skills eval. The runner gives Claude the skill and the openlabel commands as tools, runs each case against real labels, and grades the reply. A case that crashes counts as a failure, so a broken harness can’t raise a score.