What does this skill do, and when should you use it?
This skill extracts structured data from one or more clinical notes against a user-defined schema. Every populated value carries a source span and location, while missing information receives an explicit null reason. Its workflow covers schema definition, isolated per-note extraction, deterministic validation, and review reporting. It fits chart abstraction, registry work, and cohort dataset creation, but adopters must define the schema and account for terminology connectors that may be unavailable.
Converts a user's request into a schema with field descriptions, assertion axes, and optional checks; resolves pasted text, files, directories, PDFs, FHIR connectors, or database queries into plain text; runs one tool-disabled extraction worker per note; returns values, spans, locations, presence, temporality, experiencer, units, and null reasons; verifies that every populated span appears verbatim in the source; applies terminology, range, date, pattern, or enum checks; produces one row per note-field pair plus completion summaries; and can deterministically transform validated flat records into FHIR-oriented resources when FHIR is requested.
- A clinical researcher extracts predefined cohort variables from a batch of notes while retaining evidence for every populated value.
- A hospital data team builds registry fields from narrative documentation and distinguishes absent, historical, hypothetical, and non-patient assertions.
- A quality-improvement analyst organizes note evidence for care-gap or adverse-event review before applying deterministic checks.
- A clinical operations team assembles structured evidence for trial-eligibility screening or prior-authorization workflows.
- A FHIR implementation team converts validated flat records into Condition, FamilyMemberHistory, or provenance-related content.
How do you install this skill?
- Do not treat the documented evaluation table as independently reproduced accuracy evidence; the supplied material contains neither the test files nor execution results.
- Before processing real PHI, restrict source and connector permissions and define confirmation and audit rules for terminology lookups, FHIR access, and exports.
- Support for Chinese or mixed-language notes and reachability of external connectors from mainland-China networks is unproven; missing connectors leave results unvalidated.
- The skill provides extraction and validation procedures, not a substitute for clinical judgment, coding review, or safety-critical decisions.
- Shell / CLI
- Network access
- Local filesystem
- MCP Server
Bun
In Claude Code, run:
/plugin marketplace add anthropics/healthcare
/plugin install healthcare@healthcare
The repository is a healthcare plugin collection containing 11 skills. This skill is located at plugins/healthcare/skills/clinical-note-extract/SKILL.md; the source does not provide a separate installation package or procedure.
tmp="$(mktemp -d)"
git clone --depth 1 https://github.com/anthropics/healthcare.git "$tmp"
mkdir -p ~/.claude/skills
cp -R "$tmp/plugins/healthcare/skills/clinical-note-extract" ~/.claude/skills/
rm -rf "$tmp"Generated from the source repository and skill path; it copies only this skill's folder. If the author's install steps above differ, follow those first. To scope it to one project, replace ~/.claude/skills with that project's .claude/skills.
How do you use this skill?
Once installed, send your agent any of these to trigger it:
- Extract diabetes diagnosis, the latest HbA1c, and smoking status from this note; include a source span for every value and an explicit null reason for missing fields.
Define and confirm the field schema first, then provide one or more clinical notes. Example prompt: "Extract diabetes diagnosis, the latest HbA1c, and smoking status from this note; include a source span for every value and an explicit null reason for missing fields." For larger corpora, the source documents this command: bun <this skill dir>/scripts/batch.ts <notes-dir> <schema.json> records.jsonl; continue afterward with validation and reporting. Request FHIR explicitly when FHIR resources are the desired output.
What are this skill's strengths and limitations?
- Requires span-level evidence for every populated field, making results easier to audit.
- Uses explicit null reasons for missing information instead of silently treating omission as a negative finding.
- Separates extraction from deterministic checks for spans, terminology, ranges, dates, patterns, and enums.
- Supports both individual notes and batch processing, with an optional FHIR transformation.
- Requires a user-defined and confirmed schema rather than offering a fixed field set.
- Terminology validation depends on the connector named by each field's check.via; unavailable connectors leave results unvalidated.
- The inline workflow is limited to a few dozen notes per call, so larger corpora require the Bun batch script.
- The supplied material does not show a test suite, platform test results, or detailed connector configuration.
- Extracted records may contain PHI from source notes and should not be copied to shared or external systems without explicit instruction.
How does this skill compare with similar options?
Side by side with related skills; every score comes from the same FSRS standard.
| Skill | FS score | Stars | Last updated | License |
|---|---|---|---|---|
| Clinical Note Extractor this page ✓ Anthropic · Official | 56 · Use with care | ★ 425 | 1mo ago | — |
| FHIR Clinical Data Connector ✓ Anthropic · Official | 49 · Use with care | ★ 425 | 1mo ago | — |
| Lieflat Charts | 57 · Use with care | ★ 6.1k | 1mo ago | NOASSERTION |
| Data Visualization Skill | 40 · Not recommended | ★ 7.3k | 11d ago | Apache-2.0 |
| D3.js Visualization Skill | 60 · Recommended | ★ 243 | 1y ago | — |
How did FollowSkills review this skill?
The skill requires user confirmation of the schema, uses one no-tools worker per note, and clearly warns that extracted PHI remains local and must not be copied externally without authorization. Data flow and isolation are reasonably transparent. Deductions apply because it permits arbitrary tools, connectors, and source types without defining least-privilege scopes, connector trust boundaries, export confirmation, or rollback.
The steps, input contract, span validation, and deterministic checks are internally coherent, and missing connectors are explicitly reported as unvalidated. The supplied material does not include the actual workflow, runner, rules, or reproducible test files, so abnormal-input handling and diagnostic failure feedback remain thin. Under static calibration, the score cannot exceed 10.
Invocation language, target scenarios, schema inputs, record outputs, optional FHIR transformation, and batch-size limits are clearly described. Deductions apply because Chinese clinical-text support, language boundaries, and mainland-China network fit are not addressed; connector reachability is also unproven.
The documentation has clear layering and includes steps, output contracts, prerequisites, examples, extensions, limitations, and evaluation notes; publisher provenance is verified for Anthropic. Deductions apply because the skill lacks explicit versioning, changelog, maintenance ownership, and update path. License metadata is unknown, and repository-level terms do not fully establish the skill's licensing boundary.
The stated goal is to convert clinical notes into structured records with verbatim spans, assertion axes, explicit nulls, and validation status. The output contract, sample schemas, and evaluation summary support practical value. Deductions apply because the evaluation files and actual run outputs are not supplied, representative outputs are not independently verified, and clinical use still requires review.
The SKILL, README, samples, and cited clinical assertion model provide auditable primary material and disclose evaluation size and variance limitations. Deductions apply because no test suite, CI result, execution log, or independent review is supplied; the evaluation figures remain author-reported documentation. Static calibration caps this at 5.
Open a dimension to read why it scored that way
Evidence confidence：Low — Mostly static review, author material or a limited demo; useful for discovery, not high-risk decisions.
See the full review method →