FHIR Format Skill
dnlbox/fhir-format-skill
Camada de raciocínio honesta para orientação FHIR com suporte a versões e citações
Ver no GitHubSkill Claude para orientação FHIR com foco em precisão, citando fontes específicas da especificação (base, versão ou profile). Classifica afirmações por escopo, evita respostas plausíveis mas não verificadas, e integra-se com ferramentas de geração e validação de dados FHIR determinísticas para manter confiabilidade em estruturas de recursos R4/R4B/R5.
Instalação
npx convez-ai dnlbox/fhir-format-skill1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
Casos de uso
- Validar estrutura e cardinalidade de recursos FHIR com evidência de especificação
- Comparar diferenças entre versões FHIR (R4, R4B, R5) com clareza normativa
- Integrar dados FHIR em pipelines Python/TypeScript com confiabilidade verificada
- Investigar vinculações terminológicas e restrições de profile sem suposições
Veredito

Ferramenta altamente especializada para engenheiros FHIR que precisam de orientação confiável sobre padrões de interoperabilidade em saúde. Excelente em rejeitar afirmações infundadas e citar especificações, mas requer experiência prévia com FHIR e ecossistema de dados de saúde para máxima utilidade.
README
A tradução para o português deste README ainda está na fila.
fhir-format plugin skill
A Claude plugin for sourced FHIR guidance. Not a FHIR development assistant — a reasoning layer that keeps the agent honest about what the spec actually says.
Point it at any FHIR question: resource structure, cardinality rules, R4/R4B/R5 version differences, terminology bindings. The skill directs the agent to scope every claim (base spec, profile constraint, or server policy), back normative statements with citations, and flag uncertainty rather than fill gaps with plausible-sounding answers.
What makes it different from other FHIR skills: it pairs with
fhir-test-data and
fhir-resource-diff. Both tools
generate and validate FHIR resources deterministically — so when the agent reasons
about a resource, it works from a machine-verified artifact rather than recalled
spec knowledge alone.
Quick install
claude plugin marketplace add .
claude plugin install fhir-format@fhir-format-marketplace
Run the skill:
/fhir-format:fhir-format
You can also trigger it naturally with FHIR-focused prompts.
What it does
- Classifies guidance by scope: base spec, version change, profile constraint, or implementation advice
- Adds authoritative source links for base-spec statements
- Avoids guessed terminology codes and unsupported certainty
- Distinguishes structural validation results from full profile/terminology conformance
- Keeps guidance version-aware across R4, R4B, and R5
When to use
Use fhir-format for:
- Resource shape/cardinality clarification
- Version-difference checks (R4 vs R4B vs R5)
- Terminology/binding guidance with explicit evidence
- Practical integration guidance for TypeScript, Python, and AI-assisted pipelines
Avoid using fhir-format alone for:
- Final legal/compliance decisions (HIPAA/GDPR/legal counsel required)
- Full IG conformance (Implementation Guide profile conformance, e.g., US Core/IPS) sign-off without profile + terminology-capable validation
- Clinical coding decisions without official terminology lookup
Trigger examples
Natural prompts that should trigger this skill:
- "Is
Observation.coderequired in R4B?" - "Format this Patient JSON and tell me what is base FHIR vs profile-specific"
- "Can I validate this bundle with
fhir-resource-diffand call it US Core compliant?" - "What is the safest way to represent units and coding in this Observation?"
Core behavior guardrails
- No fabricated FHIR semantics
- No guessed terminology codes
- No profile constraints presented as universal base-FHIR truth
- No claim that structural checks imply full conformance
- Explicit uncertainty when source evidence is incomplete
Local development usage
From the plugin repository root:
claude --plugin-dir .
Then run:
/fhir-format:fhir-format
Toolkit workflows
# Generate + validate loop
fhir-test-data generate patient --locale uk --seed 42 | fhir-resource-diff validate - --fhir-version R4
# Fixture regression checks
fhir-test-data generate bundle --locale us --seed 1 --output ./fixtures/
fhir-resource-diff compare ./fixtures/Bundle-001.json ./baseline/ --preset metadata --exit-on-diff
# AI-friendly envelope output
fhir-test-data generate bundle --locale us | fhir-resource-diff validate - --format json --envelope
Reproducible relevance benchmark
You can independently validate whether this skill is useful and correctly scoped.
What the benchmark tests:
- Skill activation only when queries are FHIR-relevant
- Quality improvement on FHIR-focused prompts
- Better verifiability (claim labels and source-backed statements)
Runbook:
- Use
benchmark/QUERIES.mdfor the fixed query set. - Use
benchmark/PROMPTS.mdto run baseline and skill-enabled sessions. - Evaluate outputs with the evaluator prompt and store artifacts in
benchmark/results/. - Validate JSON outputs with
benchmark/validate_results.py.
See benchmark/README.md for full instructions.
Reference run artifacts are included in benchmark/results/ (dated 2026-04-08-*).
CLI grounding benchmark
There is a second benchmark lane for command-grounded tasks using fhir-test-data and fhir-resource-diff patterns.
- Query set:
benchmark/cli/QUERIES.md - Prompts:
benchmark/cli/PROMPTS.md - Runbook:
benchmark/cli/README.md - Results template:
benchmark/cli/results/RESULT_TEMPLATE.md
Reference links
- Plugins guide: https://code.claude.com/docs/en/plugins
- Plugin marketplaces guide: https://code.claude.com/docs/en/plugin-marketplaces
- Privacy policy: https://dnlbox.github.io/fhir-format-skill/privacy/