What does this skill do, and when should you use it?
This skill screens a payer’s Medicare or Medicaid claims corpus against public sources including NCCI MUE, OIG LEIE, CMS enrollment, and PFS. It runs a deterministic detection layer first, then applies model-based adjudication and provider-level narrative synthesis for SIU or program-integrity review. Findings are framed as indicators consistent with a scheme rather than proof of fraud, and cited dollar or rule allegations must trace back to independently recomputed detection results. Adoption fits teams with normalized claims data, quarter-specific reference data, and an environment that can provide local filesystem, shell, network, and workflow support.
Reads a canonical six-table corpus.duckdb plus a rule quarter and line of business; fetches and caches public reference and enrichment data when needed; runs deterministic detectors and a citation gate; adjudicates judgment-required findings by confirming, downgrading, or dismissing them; synthesizes provider narratives; and produces detection, adjudication, and final JSON snapshots, provider HTML packets, an HTML dashboard, and referrals.xlsx.
- An SIU team needs a recurring Medicare claims sweep for fraud, waste, and abuse indicators.
- A Medicaid program-integrity group needs ranked referrals from anomalous billing patterns.
- A payer needs findings checked against NCCI MUE, LEIE, CMS enrollment, and PFS rules with reproducible citations.
- Investigators need provider-organized review packets with source excerpts and cited rule allegations.
How do you install this skill?
- Do not treat the output as a fraud determination; SIU or compliance investigators must review it.
- Before supplying a claims database, connection string, or FHIR credential, confirm read-only access, credential handling, data minimization, and access auditing.
- Reference fetching depends on CMS, OIG, and state-policy sites; unreachable overseas networks, URL changes, or bot protection may produce incomplete results.
- Before running, verify the quarter's reference database, DuckDB schema, Node.js, unzip, pdftotext, Workflow, and Write tools, and retain failure logs.
- Shell / CLI
- Network access
- Local filesystem
- MCP Server
Node.jsunzippdftotext (poppler)
Add the repository to the Claude Code plugin marketplace and install the healthcare plugin: /plugin marketplace add anthropics/healthcare, followed by /plugin install healthcare@healthcare. The README documents installation for the full healthcare collection, not a separate installation command for this skill.
tmp="$(mktemp -d)"
git clone --depth 1 https://github.com/anthropics/healthcare.git "$tmp"
mkdir -p ~/.claude/skills
cp -R "$tmp/plugins/healthcare/skills/fraud-detection" ~/.claude/skills/
rm -rf "$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:
- Run a fraud sweep on this Medicaid claims corpus using the 2026q3 rules and generate investigation referrals.
Provide a payer claims corpus.duckdb with the documented canonical six-table schema, a reference quarter, and either medicare or medicaid. Example trigger: “Run a fraud sweep on this Medicaid claims corpus using the 2026q3 rules and generate investigation referrals.” The client must provide the documented Workflow, shell, network, and local filesystem capabilities.
What are this skill's strengths and limitations?
- Separates deterministic calculations from model adjudication and narrative synthesis, with citations gated against recomputation.
- Supports Medicare and Medicaid and references NCCI MUE, OIG LEIE, CMS enrollment, PFS, and local enrichment data.
- Produces ranked referrals, provider review packets, an HTML dashboard, and an XLSX export.
- Uses cautious SIU language and does not treat pattern matches as established fraud.
- Requires a payer claims corpus.duckdb in the specified six-table schema.
- Initial or new-quarter reference seeding requires network access and depends on Node.js, unzip, and pdftotext.
- The Workflow tool and /workflows experience are Claude Code-specific and require adaptation elsewhere.
- The supplied materials show no test suite, performance benchmarks, or supported corpus-size limits.
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 |
|---|---|---|---|---|
| Healthcare Claims Fraud Screening this page ✓ Anthropic · Official | 48 · Use with care | ★ 425 | 1mo ago | — |
| Hugging Face Datasets Skill | 50 · Use with care | ★ 38k | 13d ago | MIT |
| Azure Data Explorer Kusto Analytics ✓ Microsoft · Official | 45 · Not recommended | ★ 2.7k | 3mo ago | MIT |
| Workspace Data Analyst | 54 · Use with care | ★ 11k | 27d ago | MIT |
| Claude Ecom — Ecommerce Business Review | 53 · Use with care | ★ 53 | 4mo ago | MIT |
How did FollowSkills review this skill?
The evidence shows deterministic detection, a citation gate, human SIU review, no autonomous publishing, and data stored outside the install path; LOAD-CLAIMS also requests read-only access. However, the skill handles highly sensitive medical claims and identity data while requesting connection strings, credentials, or FHIR bearer tokens without sufficient detail on credential isolation, field minimization, access auditing, encryption, deletion, or rollback, so points are deducted.
The documentation provides a fairly complete procedure, directory conventions, stage snapshots, and some recovery guidance, and separates detection, adjudication, and synthesis. Static evidence does not establish availability of the key scripts, Workflow and Write tools, or external dependencies in the target environment; unified diagnostics for network failures, missing tables, tool failures, and abnormal inputs are also limited, so the score remains conservative under the static cap.
The audience, corpus input, quarter, line of business, and outputs are reasonably clear, and the non-fit boundaries for real-time editing and final fraud determinations are stated. Trigger exclusions, payer data-quality boundaries, Chinese-language operation, and mainland-China network reachability are not specified; core reference fetching depends on CMS, OIG, and other overseas services, which may limit access for FollowSkills users.
The documentation is well organized and includes quickstart material, claim loading, reference data, architecture, limitations, licensing notes, and extension guidance; official organizational provenance is supplied. The skill lacks clear versioning, changelog, named maintenance responsibility, or an update commitment, while license metadata is unknown and dependency/tool prerequisites are incomplete, so points are deducted.
The intended task and output format are clear, and deterministic screening, cited referral packets, dashboards, and XLSX outputs could be directly useful; the documentation also requires dollar and rule claims to trace to the detection stage. Because this is a static review with no representative run results, sample outputs, or real-data coverage evidence, end-to-end completeness and boundary quality cannot be confirmed, so the score is capped at 7.
The materials describe deterministic detectors, a citation gate, stage snapshots, cohort data for recomputation, and repository CI-related files. They do not provide executed results for this skill's key paths, proof of committed test coverage, or independent review records, so the score is limited to the strongest level supported by static evidence.
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 →