[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"skill-aws-labs-aws-genai-ml-architect":3,"mdc-kxhb3k-key":52,"related-org-aws-labs-aws-genai-ml-architect":1290,"related-repo-aws-labs-aws-genai-ml-architect":1468},{"slug":4,"name":4,"fn":5,"description":6,"org":7,"tags":12,"stars":29,"repoUrl":30,"updatedAt":31,"license":32,"forks":33,"topics":34,"repo":47,"sourceUrl":50,"mdContent":51},"aws-genai-ml-architect","design AWS GenAI and ML architectures","Reasoning skill for designing AWS GenAI and ML architectures for healthcare and life sciences workloads. Use when the user asks to choose between SageMaker and Bedrock, design a RAG system over medical literature, architect clinical NLP or medical imaging inference, plan genomics or drug discovery pipelines on AWS, address HIPAA\u002FPHI compliance in ML systems, design MLOps for regulated clinical models, or optimize cost for HCLS ML workloads. Triggers include \"AWS architecture\", \"SageMaker vs Bedrock\", \"HIPAA ML\", \"clinical RAG\", \"medical imaging inference\", \"genomics on AWS\", \"PHI training\", \"MLOps healthcare\", \"Bedrock guardrails\", \"HealthLake\", \"HCLS cloud architecture\", \"BAA compliance\", \"SageMaker endpoint\", \"Bedrock knowledge base\", \"clinical NLP on AWS\", \"FDA SaMD on AWS\".",{"slug":8,"name":9,"logoUrl":10,"githubOrg":11},"aws-labs","AWS Labs","https:\u002F\u002Fpexgzepcugksgbtrxkhf.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Forg-logos\u002Faws-labs.png","awslabs",[13,17,20,23,26],{"name":14,"slug":15,"type":16},"LLM","llm","tag",{"name":18,"slug":19,"type":16},"Architecture","architecture",{"name":21,"slug":22,"type":16},"Healthcare","healthcare",{"name":24,"slug":25,"type":16},"Life Sciences","life-sciences",{"name":27,"slug":28,"type":16},"AWS","aws",4,"https:\u002F\u002Fgithub.com\u002Fawslabs\u002Fhcls-agent-skills","2026-07-12T08:38:07.975937",null,0,[35,36,37,38,39,40,41,42,43,25,44,45,46],"agent-skills","agentcore","ai-agents","amazon-quick-desktop","claims-processing","drug-discovery","genomics","healthcare-ai","kiro","medical-imaging","risk-adjustment","strands-agents",{"repoUrl":30,"stars":29,"forks":33,"topics":48,"description":49},[35,36,37,38,39,40,41,42,43,25,44,45,46],"Agent skills for healthcare and life sciences: genomics, imaging, claims, drug discovery, and more. Works with Amazon Quick, Kiro, Amazon AgentCore, AWS Strands SDK, Claude Code, Codex, and any Agent Skills-compatible platform.","https:\u002F\u002Fgithub.com\u002Fawslabs\u002Fhcls-agent-skills\u002Ftree\u002FHEAD\u002Fskills\u002Faws-genai-ml-architect","---\nname: aws-genai-ml-architect\ndescription: Reasoning skill for designing AWS GenAI and ML architectures for healthcare and life sciences workloads. Use when the user asks to choose between SageMaker and Bedrock, design a RAG system over medical literature, architect clinical NLP or medical imaging inference, plan genomics or drug discovery pipelines on AWS, address HIPAA\u002FPHI compliance in ML systems, design MLOps for regulated clinical models, or optimize cost for HCLS ML workloads. Triggers include \"AWS architecture\", \"SageMaker vs Bedrock\", \"HIPAA ML\", \"clinical RAG\", \"medical imaging inference\", \"genomics on AWS\", \"PHI training\", \"MLOps healthcare\", \"Bedrock guardrails\", \"HealthLake\", \"HCLS cloud architecture\", \"BAA compliance\", \"SageMaker endpoint\", \"Bedrock knowledge base\", \"clinical NLP on AWS\", \"FDA SaMD on AWS\".\nusage: Invoke when designing AWS GenAI\u002FML architectures for HCLS workloads, choosing between SageMaker and Bedrock, or addressing HIPAA compliance.\nversion: 1.0.0\ntags: [skill, category:reasoning, aws, sagemaker, bedrock, architecture, hcls, genai]\n---\n\n# AWS GenAI\u002FML Architect (HCLS)\n\n## Overview\n\nThis skill teaches the agent how to reason about AWS GenAI and ML architecture\ndecisions for healthcare and life sciences (HCLS) workloads. It does not\nprescribe CLI commands or IaC templates — it encodes the decision frameworks\nan experienced HCLS cloud architect uses to pick services, enforce compliance,\nand avoid common pitfalls. Pair this skill with a pipeline skill when concrete IaC or deployment artifacts are needed.\n\nHCLS workloads are distinguished from generic ML workloads by three constraints\nthat dominate every decision: PHI handling under HIPAA, high cost of\nhallucination in clinical outputs, and distribution shift driven by coding\nchanges, site onboarding, and seasonal care patterns. Keep these in mind at\nevery step.\n\n## Usage\n\nInvoke this skill when the user asks architecture-level questions such as:\n\n- \"Should I use Bedrock or SageMaker for summarizing clinical notes?\"\n- \"How do I build RAG over medical literature on AWS?\"\n- \"What's the right inference pattern for whole-slide images?\"\n- \"How do I train a model on PHI safely?\"\n- \"What should my MLOps stack look like for a clinical risk model?\"\n\nWork through the decisions in this order:\n\n1. Clarify the clinical \u002F scientific use case and PHI exposure.\n2. Apply the service selection framework.\n3. Overlay HCLS architecture patterns for the specific modality.\n4. Enforce security and compliance decision points.\n5. Add MLOps, monitoring, and cost considerations.\n6. Surface common mistakes before finalizing.\n\nState assumptions explicitly when the user has not specified payload size,\nlatency SLO, PHI status, or regulatory posture.\n\n## Response Format\n\n- Lead with the direct recommendation or classification (≤3 sentences)\n- Structure as: recommendation → justification (citing specific criteria\u002Fthresholds) → caveats\n- Use tables for comparisons; bullet points for criteria lists\n- Omit background the user already knows — they asked the question\n- Target: 200-400 words unless the user requests exhaustive detail\nThe decision trees and frameworks in this skill are for internal reasoning only. Apply them to reach your conclusion, but do not reproduce them in your response. Present only the final recommendation with supporting evidence.\n\n\n## Core Concepts\n\n- **PHI vs. de-identified data**: PHI requires HIPAA-eligible services, a\n  signed BAA, and stricter isolation. De-identified or synthetic data relaxes\n  some controls but downstream re-identification risk should still be assessed.\n- **Foundation model (FM) vs. custom model**: FMs are general, fast to ship,\n  and priced per token\u002Fcall. Custom models are domain-specific, require\n  training data and ops, and are priced per compute hour.\n- **Sync vs. async vs. batch inference**: Latency SLO, payload size, and\n  traffic shape determine the right endpoint type, not model size alone.\n- **Grounding vs. generation**: Clinical outputs must be grounded in\n  retrievable sources (RAG, structured lookups) whenever a wrong answer can\n  harm patients or invite regulatory risk.\n- **Distribution shift is the default, not the exception**: Healthcare data\n  shifts with every ICD coding update, new site, season, and care pathway\n  change. Design for it from day one.\n\n## Service Selection Framework\n\nUse this as the first-cut decision:\n\n- **Bedrock** when:\n  - Using a foundation model as-is (summarization, extraction, Q&A over text)\n  - Building RAG with Knowledge Bases or Agents\n  - You want managed guardrails for PII\u002FPHI redaction and toxicity\n  - You do not want to manage training, endpoints, or GPU capacity\n\n- **SageMaker** when:\n  - Custom training or fine-tuning on proprietary clinical or -omics data\n  - Specialized models (medical imaging CNNs, custom NER, survival models)\n  - You need full control over container, network, and instance type\n  - Building MLOps pipelines with model registry and approval workflows\n\n- **Both** when:\n  - FM handles unstructured text while a custom model scores structured risk\n  - An ensemble combines an FM summary with a specialized classifier\n  - RAG retrieval uses a custom embedding model trained on clinical text\n\nRule of thumb: start with Bedrock for any text-centric use case. Move to\nSageMaker only when FM accuracy, cost, latency, or data residency forces it.\n\n## HCLS Architecture Patterns\n\nMatch the modality to the pattern:\n\n- **Clinical note summarization \u002F abstraction**: Bedrock with a small,\n  instruction-tuned FM, prompt templates reviewed by clinicians, and\n  Bedrock Guardrails configured for PHI redaction and denied topics.\n  Always require source-span citations back into the note.\n\n- **Medical literature \u002F guideline RAG**: Bedrock Knowledge Bases backed by\n  OpenSearch Serverless. Chunk by semantic section (abstract, methods,\n  results, guideline recommendation) rather than fixed token windows —\n  clinicians reason at section granularity, and fixed chunks split tables\n  and dosing guidance. Store source metadata (PMID, guideline version,\n  publication date) for traceability.\n\n- **Pathology and radiology inference**: SageMaker async endpoints for\n  whole-slide images and DICOM volumes. Payloads exceed real-time endpoint\n  limits and inference times exceed sync SLOs. Pre-process (tiling,\n  resampling) in SageMaker Processing jobs before invoking the endpoint.\n\n- **Genomics pipelines**: SageMaker Processing jobs orchestrated by Step\n  Functions for custom workflows; HealthOmics when the workflow fits its\n  managed WDL\u002FNextflow execution model. Use HealthOmics variant and\n  annotation stores for queryable genomic data.\n\n- **Drug discovery and molecular modeling**: SageMaker HyperPod for\n  distributed training of large structural or property-prediction models.\n  Bedrock for molecular description, literature mining, and chemistry Q&A;\n  never for final property prediction where accuracy is load-bearing.\n\n- **Ambient clinical documentation**: Bedrock for transcription\n  post-processing and structured note generation; guardrails plus\n  human-in-the-loop sign-off in the EHR.\n\n## Data Architecture\n\nPick the store by data shape and access pattern:\n\n- **FHIR clinical data** → HealthLake. Supports FHIR-native queries and\n  integrated NLP. Export to S3 for ML training.\n- **Observational \u002F claims research** → OMOP CDM on Redshift or Athena.\n  Standard vocabulary enables reuse of published phenotype definitions.\n- **Genomics** → S3 as the source of truth for VCF, BAM, CRAM, FASTQ;\n  Lake Formation for governance; Athena for ad-hoc queries; HealthOmics\n  stores for indexed access.\n- **Medical imaging** → S3 for raw DICOM; HealthImaging for indexed,\n  sub-image access and integration with inference pipelines.\n- **Unstructured documents** (PDFs, scanned notes) → S3 + Textract for\n  extraction, then route structured output to the appropriate store above.\n\nCross-cutting: Lake Formation for fine-grained access control across\nanalyst personas; Glue Data Catalog as the single metadata plane.\n\n## Security & Compliance\n\nTreat compliance as a set of hard gates, not a checklist at the end.\n\n- **HIPAA BAA**: Only use services on the AWS HIPAA-eligible list for any\n  workflow touching PHI. Verify eligibility for every service in the\n  architecture, including logging and monitoring sinks.\n- **Network isolation**: PHI workloads run in private subnets with no\n  internet egress. Use VPC endpoints for S3, SageMaker, Bedrock, KMS, and\n  CloudWatch. Enable SageMaker network isolation on training jobs and\n  endpoints handling PHI.\n- **Encryption**: Customer-managed KMS keys for S3, EBS, SageMaker\n  volumes, and Bedrock Knowledge Base vector stores. TLS 1.2 or higher in\n  transit. Scope key policies so only the intended roles can decrypt.\n- **Audit and detection**: CloudTrail (including data events for PHI\n  buckets), AWS Config rules for drift, GuardDuty for threat detection.\n  Enable Bedrock model invocation logging to S3 or CloudWatch — without it\n  you cannot reconstruct what the FM was asked or returned.\n- **IAM**: Least privilege, no wildcards on PHI buckets or KMS keys.\n  Separate roles for data scientists, training jobs, and inference\n  endpoints. Never share a role across environments.\n- **De-identification**: If the use case permits, de-identify before\n  training. Document the method (Safe Harbor, Expert Determination) and\n  retain the determination evidence.\n\nCompliance decision points to surface explicitly:\n\n- Is any input, output, prompt, or log PHI? If yes, every service in the\n  path must be HIPAA-eligible and under BAA.\n- Will the FM provider see PHI? Confirm the Bedrock model is in an\n  eligible region and that invocation logs are encrypted with your CMK.\n- Does the workload cross accounts or regions? Re-verify BAA coverage and\n  data residency commitments.\n\n## MLOps for HCLS\n\nClinical models decay. Assume drift and design for it:\n\n- **Monitoring**: SageMaker Model Monitor for data quality, model quality\n  (when labels arrive), and bias drift. Track feature distributions at the\n  site and payer level, not only in aggregate.\n- **Distribution shift sources to watch**: ICD and CPT coding updates,\n  EHR template changes, new site onboarding, seasonal disease patterns,\n  formulary changes, care-pathway interventions.\n- **Model registry and approvals**: Every clinical model goes through the\n  SageMaker Model Registry with explicit approval states. Gate promotion\n  on model card completion, bias evaluation, and clinical sign-off.\n- **Experimentation constraints**: Randomized online A\u002FB testing on\n  patients typically requires IRB approval. Default to shadow mode\n  evaluation or champion\u002Fchallenger with human-in-the-loop review.\n  Document the evaluation design before enabling traffic shifts.\n- **Reproducibility**: Pin container images, training data snapshots\n  (S3 versioning), and hyperparameters in the registry. A model that\n  cannot be rebuilt cannot be defended to regulators.\n- **Incident response**: Define rollback criteria and a kill switch for\n  every deployed clinical model before go-live.\n\n## Cost Optimization\n\nOptimize only after correctness and compliance are settled:\n\n- **Training**: Spot instances for fault-tolerant training; checkpointing\n  is mandatory — without it spot interruptions waste compute. HyperPod for\n  long-running distributed jobs to amortize setup.\n- **Inference shape**:\n  - Real-time endpoint: latency \u003C ~60s and payload \u003C ~6 MB with steady\n    traffic.\n  - Async endpoint: large payloads (WSI, DICOM volumes, long documents)\n    or long inference times.\n  - Serverless inference: bursty or intermittent traffic with cold-start\n    tolerance.\n  - Batch transform: offline scoring of large cohorts.\n- **Storage**: S3 Intelligent-Tiering for research datasets with unknown\n  access patterns; lifecycle policies to Glacier for archival imaging and\n  genomics data retained for compliance.\n- **Bedrock**: Prefer smaller models for extraction and classification;\n  reserve larger models for open-ended generation. Cache retrieval results\n  where clinically safe.\n\n## Decision Trees\n\nWalk the user through these when the path is not obvious.\n\n**Processing clinical text**\n1. Is the task extraction, summarization, or Q&A over text? → Bedrock.\n2. Is it a specialized task (custom NER schema, de-identification at\n   scale, domain-specific classifier)? → SageMaker with a fine-tuned\n   clinical model.\n3. Is PHI involved? → Apply the Security & Compliance gates regardless of\n   which service is chosen.\n\n**Training on PHI**\n1. SageMaker training job in a VPC, private subnets only.\n2. Network isolation enabled on the job.\n3. KMS CMK on input S3, training volume, and output artifacts.\n4. BAA-covered region and services end to end.\n5. No internet egress; VPC endpoints for all AWS APIs used.\n\n**Choosing an inference pattern**\n1. Latency SLO under a few seconds and payload small? → Real-time endpoint.\n2. Payload large (WSI, DICOM, long report) or inference slow? → Async\n   endpoint with S3 input and output.\n3. Traffic bursty or low-volume? → Serverless inference.\n4. Scoring a full cohort offline? → Batch transform.\n\n**RAG over clinical or scientific content**\n1. Is the content stable and well-structured? → Bedrock Knowledge Bases\n   with OpenSearch Serverless.\n2. Chunk by section, not fixed tokens. Preserve tables and dosing blocks.\n3. Require citations in every answer; reject ungrounded responses.\n4. If the corpus includes PHI, ensure the vector store, embedding model,\n   and FM are all HIPAA-eligible and under BAA.\n\n## When NOT to Use This Skill\n- Non-HCLS workloads — this skill's compliance guidance (HIPAA, BAA, FDA SaMD) does not apply to general-purpose applications\n- Cost estimation or pricing — use the AWS Pricing Calculator; this skill covers architecture patterns, not billing\n- Reviewing or auditing an existing deployed architecture — this skill designs new systems; use Well-Architected Reviews for existing ones\n\n## Common Mistakes\n\n- **Wrong:** Using Bedrock for a problem that needs a custom model\n  **Right:** Use SageMaker with a fine-tuned classifier when accuracy is load-bearing and the label space is narrow\n  **Why:** FMs underperform well-labeled domain classifiers on constrained tasks, producing unreliable clinical outputs\n\n- **Wrong:** Using SageMaker when Bedrock would ship in a week\n  **Right:** Start with Bedrock for generic summarization or extraction tasks; move to SageMaker only when forced by accuracy, cost, or data residency\n  **Why:** Building bespoke infrastructure for a commodity NLP task wastes months of engineering time\n\n- **Wrong:** Fixed-size chunking for medical literature in RAG pipelines\n  **Right:** Chunk by semantic section (abstract, methods, results, guideline recommendation) and preserve tables and dosing blocks intact\n  **Why:** Fixed chunks split dosing tables, contraindications, and guideline recommendations, destroying retrieval quality\n\n- **Wrong:** Using real-time endpoints for whole-slide images\n  **Right:** Use SageMaker async endpoints for WSI and DICOM volumes\n  **Why:** WSI payloads exceed real-time endpoint size limits and inference times exceed sync timeout SLOs\n\n- **Wrong:** Forgetting to enable Bedrock invocation logging\n  **Right:** Always enable Bedrock model invocation logging to S3 or CloudWatch with CMK encryption\n  **Why:** Without it you cannot reconstruct what the FM was asked or returned — a compliance and debugging failure\n\n- **Wrong:** Using public subnets or internet egress in PHI training jobs\n  **Right:** Run PHI workloads in private subnets with no internet egress; use VPC endpoints for all AWS APIs\n  **Why:** Violates HIPAA isolation expectations even if data itself is encrypted in transit\n\n- **Wrong:** Using wildcard IAM policies on PHI buckets\n  **Right:** Scope IAM policies to specific resources and actions; separate roles per environment and persona\n  **Why:** A single over-scoped role undermines every other security control in the architecture\n\n- **Wrong:** Training on spot instances without checkpointing\n  **Right:** Always enable checkpointing when using spot instances for training jobs\n  **Why:** Spot interruptions lose all progress since last checkpoint; often ends up more expensive than on-demand\n\n- **Wrong:** Running randomized online A\u002FB tests on patients without IRB approval\n  **Right:** Default to shadow mode evaluation or champion\u002Fchallenger with human-in-the-loop review\n  **Why:** Randomized experiments on patients constitute human subjects research — a regulatory violation without IRB\n\n- **Wrong:** Ignoring distribution shift until aggregate accuracy drops\n  **Right:** Monitor feature distributions at the site and payer level from day one using SageMaker Model Monitor\n  **Why:** By the time aggregate metrics move, site-level harm may already have occurred\n\n- **Wrong:** Assuming de-identification removes all data handling obligations\n  **Right:** Document the de-identification method (Safe Harbor or Expert Determination) and assess re-identification risk and contractual terms\n  **Why:** Re-identification risk and contractual obligations often persist after de-identification\n\n- **Wrong:** Mixing dev and prod in one AWS account with shared KMS keys\n  **Right:** Separate dev and prod into distinct accounts with independent KMS keys and IAM boundaries\n  **Why:** Blast radius of any mistake in the shared account covers production PHI\n\n## When to Escalate to a Human Expert\n\nRecommend human review when:\n\n- The use case directly drives a treatment decision without clinician\n  review in the loop.\n- The workload requires FDA SaMD classification or 21 CFR Part 11 controls\n  you cannot confirm.\n- Data residency, cross-border transfer, or multi-jurisdiction PHI rules\n  apply.\n- A novel FM is being proposed for a safety-critical clinical output with\n  no published evaluation on the target population.\n\nIn these cases, draft the architecture, list the open regulatory and\nclinical questions, and stop before implementation.\n\n## References\n\n- HIPAA eligible AWS services: https:\u002F\u002Faws.amazon.com\u002Fcompliance\u002Fhipaa-eligible-services-reference\u002F\n- AWS Well-Architected ML Lens: https:\u002F\u002Fdocs.aws.amazon.com\u002Fwellarchitected\u002Flatest\u002Fmachine-learning-lens\u002Fmachine-learning-lens.html\n- SageMaker documentation: https:\u002F\u002Fdocs.aws.amazon.com\u002Fsagemaker\u002F\n- Bedrock documentation: https:\u002F\u002Fdocs.aws.amazon.com\u002Fbedrock\u002F\n",{"data":53,"body":63},{"name":4,"description":6,"usage":54,"version":55,"tags":56},"Invoke when designing AWS GenAI\u002FML architectures for HCLS workloads, choosing between SageMaker and Bedrock, or addressing HIPAA compliance.","1.0.0",[57,58,28,59,60,19,61,62],"skill","category:reasoning","sagemaker","bedrock","hcls","genai",{"type":64,"children":65},"root",[66,75,82,88,93,99,104,134,139,173,178,184,212,218,272,278,283,378,383,389,394,457,463,468,521,526,532,537,600,605,623,629,634,697,703,708,773,779,784,792,810,818,846,854,877,885,908,914,932,938,1196,1202,1207,1230,1235,1241],{"type":67,"tag":68,"props":69,"children":71},"element","h1",{"id":70},"aws-genaiml-architect-hcls",[72],{"type":73,"value":74},"text","AWS GenAI\u002FML Architect (HCLS)",{"type":67,"tag":76,"props":77,"children":79},"h2",{"id":78},"overview",[80],{"type":73,"value":81},"Overview",{"type":67,"tag":83,"props":84,"children":85},"p",{},[86],{"type":73,"value":87},"This skill teaches the agent how to reason about AWS GenAI and ML architecture\ndecisions for healthcare and life sciences (HCLS) workloads. It does not\nprescribe CLI commands or IaC templates — it encodes the decision frameworks\nan experienced HCLS cloud architect uses to pick services, enforce compliance,\nand avoid common pitfalls. Pair this skill with a pipeline skill when concrete IaC or deployment artifacts are needed.",{"type":67,"tag":83,"props":89,"children":90},{},[91],{"type":73,"value":92},"HCLS workloads are distinguished from generic ML workloads by three constraints\nthat dominate every decision: PHI handling under HIPAA, high cost of\nhallucination in clinical outputs, and distribution shift driven by coding\nchanges, site onboarding, and seasonal care patterns. Keep these in mind at\nevery step.",{"type":67,"tag":76,"props":94,"children":96},{"id":95},"usage",[97],{"type":73,"value":98},"Usage",{"type":67,"tag":83,"props":100,"children":101},{},[102],{"type":73,"value":103},"Invoke this skill when the user asks architecture-level questions such as:",{"type":67,"tag":105,"props":106,"children":107},"ul",{},[108,114,119,124,129],{"type":67,"tag":109,"props":110,"children":111},"li",{},[112],{"type":73,"value":113},"\"Should I use Bedrock or SageMaker for summarizing clinical notes?\"",{"type":67,"tag":109,"props":115,"children":116},{},[117],{"type":73,"value":118},"\"How do I build RAG over medical literature on AWS?\"",{"type":67,"tag":109,"props":120,"children":121},{},[122],{"type":73,"value":123},"\"What's the right inference pattern for whole-slide images?\"",{"type":67,"tag":109,"props":125,"children":126},{},[127],{"type":73,"value":128},"\"How do I train a model on PHI safely?\"",{"type":67,"tag":109,"props":130,"children":131},{},[132],{"type":73,"value":133},"\"What should my MLOps stack look like for a clinical risk model?\"",{"type":67,"tag":83,"props":135,"children":136},{},[137],{"type":73,"value":138},"Work through the decisions in this order:",{"type":67,"tag":140,"props":141,"children":142},"ol",{},[143,148,153,158,163,168],{"type":67,"tag":109,"props":144,"children":145},{},[146],{"type":73,"value":147},"Clarify the clinical \u002F scientific use case and PHI exposure.",{"type":67,"tag":109,"props":149,"children":150},{},[151],{"type":73,"value":152},"Apply the service selection framework.",{"type":67,"tag":109,"props":154,"children":155},{},[156],{"type":73,"value":157},"Overlay HCLS architecture patterns for the specific modality.",{"type":67,"tag":109,"props":159,"children":160},{},[161],{"type":73,"value":162},"Enforce security and compliance decision points.",{"type":67,"tag":109,"props":164,"children":165},{},[166],{"type":73,"value":167},"Add MLOps, monitoring, and cost considerations.",{"type":67,"tag":109,"props":169,"children":170},{},[171],{"type":73,"value":172},"Surface common mistakes before finalizing.",{"type":67,"tag":83,"props":174,"children":175},{},[176],{"type":73,"value":177},"State assumptions explicitly when the user has not specified payload size,\nlatency SLO, PHI status, or regulatory posture.",{"type":67,"tag":76,"props":179,"children":181},{"id":180},"response-format",[182],{"type":73,"value":183},"Response Format",{"type":67,"tag":105,"props":185,"children":186},{},[187,192,197,202,207],{"type":67,"tag":109,"props":188,"children":189},{},[190],{"type":73,"value":191},"Lead with the direct recommendation or classification (≤3 sentences)",{"type":67,"tag":109,"props":193,"children":194},{},[195],{"type":73,"value":196},"Structure as: recommendation → justification (citing specific criteria\u002Fthresholds) → caveats",{"type":67,"tag":109,"props":198,"children":199},{},[200],{"type":73,"value":201},"Use tables for comparisons; bullet points for criteria lists",{"type":67,"tag":109,"props":203,"children":204},{},[205],{"type":73,"value":206},"Omit background the user already knows — they asked the question",{"type":67,"tag":109,"props":208,"children":209},{},[210],{"type":73,"value":211},"Target: 200-400 words unless the user requests exhaustive detail\nThe decision trees and frameworks in this skill are for internal reasoning only. Apply them to reach your conclusion, but do not reproduce them in your response. Present only the final recommendation with supporting evidence.",{"type":67,"tag":76,"props":213,"children":215},{"id":214},"core-concepts",[216],{"type":73,"value":217},"Core Concepts",{"type":67,"tag":105,"props":219,"children":220},{},[221,232,242,252,262],{"type":67,"tag":109,"props":222,"children":223},{},[224,230],{"type":67,"tag":225,"props":226,"children":227},"strong",{},[228],{"type":73,"value":229},"PHI vs. de-identified data",{"type":73,"value":231},": PHI requires HIPAA-eligible services, a\nsigned BAA, and stricter isolation. De-identified or synthetic data relaxes\nsome controls but downstream re-identification risk should still be assessed.",{"type":67,"tag":109,"props":233,"children":234},{},[235,240],{"type":67,"tag":225,"props":236,"children":237},{},[238],{"type":73,"value":239},"Foundation model (FM) vs. custom model",{"type":73,"value":241},": FMs are general, fast to ship,\nand priced per token\u002Fcall. Custom models are domain-specific, require\ntraining data and ops, and are priced per compute hour.",{"type":67,"tag":109,"props":243,"children":244},{},[245,250],{"type":67,"tag":225,"props":246,"children":247},{},[248],{"type":73,"value":249},"Sync vs. async vs. batch inference",{"type":73,"value":251},": Latency SLO, payload size, and\ntraffic shape determine the right endpoint type, not model size alone.",{"type":67,"tag":109,"props":253,"children":254},{},[255,260],{"type":67,"tag":225,"props":256,"children":257},{},[258],{"type":73,"value":259},"Grounding vs. generation",{"type":73,"value":261},": Clinical outputs must be grounded in\nretrievable sources (RAG, structured lookups) whenever a wrong answer can\nharm patients or invite regulatory risk.",{"type":67,"tag":109,"props":263,"children":264},{},[265,270],{"type":67,"tag":225,"props":266,"children":267},{},[268],{"type":73,"value":269},"Distribution shift is the default, not the exception",{"type":73,"value":271},": Healthcare data\nshifts with every ICD coding update, new site, season, and care pathway\nchange. Design for it from day one.",{"type":67,"tag":76,"props":273,"children":275},{"id":274},"service-selection-framework",[276],{"type":73,"value":277},"Service Selection Framework",{"type":67,"tag":83,"props":279,"children":280},{},[281],{"type":73,"value":282},"Use this as the first-cut decision:",{"type":67,"tag":105,"props":284,"children":285},{},[286,319,351],{"type":67,"tag":109,"props":287,"children":288},{},[289,294,296],{"type":67,"tag":225,"props":290,"children":291},{},[292],{"type":73,"value":293},"Bedrock",{"type":73,"value":295}," when:",{"type":67,"tag":105,"props":297,"children":298},{},[299,304,309,314],{"type":67,"tag":109,"props":300,"children":301},{},[302],{"type":73,"value":303},"Using a foundation model as-is (summarization, extraction, Q&A over text)",{"type":67,"tag":109,"props":305,"children":306},{},[307],{"type":73,"value":308},"Building RAG with Knowledge Bases or Agents",{"type":67,"tag":109,"props":310,"children":311},{},[312],{"type":73,"value":313},"You want managed guardrails for PII\u002FPHI redaction and toxicity",{"type":67,"tag":109,"props":315,"children":316},{},[317],{"type":73,"value":318},"You do not want to manage training, endpoints, or GPU capacity",{"type":67,"tag":109,"props":320,"children":321},{},[322,327,328],{"type":67,"tag":225,"props":323,"children":324},{},[325],{"type":73,"value":326},"SageMaker",{"type":73,"value":295},{"type":67,"tag":105,"props":329,"children":330},{},[331,336,341,346],{"type":67,"tag":109,"props":332,"children":333},{},[334],{"type":73,"value":335},"Custom training or fine-tuning on proprietary clinical or -omics data",{"type":67,"tag":109,"props":337,"children":338},{},[339],{"type":73,"value":340},"Specialized models (medical imaging CNNs, custom NER, survival models)",{"type":67,"tag":109,"props":342,"children":343},{},[344],{"type":73,"value":345},"You need full control over container, network, and instance type",{"type":67,"tag":109,"props":347,"children":348},{},[349],{"type":73,"value":350},"Building MLOps pipelines with model registry and approval workflows",{"type":67,"tag":109,"props":352,"children":353},{},[354,359,360],{"type":67,"tag":225,"props":355,"children":356},{},[357],{"type":73,"value":358},"Both",{"type":73,"value":295},{"type":67,"tag":105,"props":361,"children":362},{},[363,368,373],{"type":67,"tag":109,"props":364,"children":365},{},[366],{"type":73,"value":367},"FM handles unstructured text while a custom model scores structured risk",{"type":67,"tag":109,"props":369,"children":370},{},[371],{"type":73,"value":372},"An ensemble combines an FM summary with a specialized classifier",{"type":67,"tag":109,"props":374,"children":375},{},[376],{"type":73,"value":377},"RAG retrieval uses a custom embedding model trained on clinical text",{"type":67,"tag":83,"props":379,"children":380},{},[381],{"type":73,"value":382},"Rule of thumb: start with Bedrock for any text-centric use case. Move to\nSageMaker only when FM accuracy, cost, latency, or data residency forces it.",{"type":67,"tag":76,"props":384,"children":386},{"id":385},"hcls-architecture-patterns",[387],{"type":73,"value":388},"HCLS Architecture Patterns",{"type":67,"tag":83,"props":390,"children":391},{},[392],{"type":73,"value":393},"Match the modality to the pattern:",{"type":67,"tag":105,"props":395,"children":396},{},[397,407,417,427,437,447],{"type":67,"tag":109,"props":398,"children":399},{},[400,405],{"type":67,"tag":225,"props":401,"children":402},{},[403],{"type":73,"value":404},"Clinical note summarization \u002F abstraction",{"type":73,"value":406},": Bedrock with a small,\ninstruction-tuned FM, prompt templates reviewed by clinicians, and\nBedrock Guardrails configured for PHI redaction and denied topics.\nAlways require source-span citations back into the note.",{"type":67,"tag":109,"props":408,"children":409},{},[410,415],{"type":67,"tag":225,"props":411,"children":412},{},[413],{"type":73,"value":414},"Medical literature \u002F guideline RAG",{"type":73,"value":416},": Bedrock Knowledge Bases backed by\nOpenSearch Serverless. Chunk by semantic section (abstract, methods,\nresults, guideline recommendation) rather than fixed token windows —\nclinicians reason at section granularity, and fixed chunks split tables\nand dosing guidance. Store source metadata (PMID, guideline version,\npublication date) for traceability.",{"type":67,"tag":109,"props":418,"children":419},{},[420,425],{"type":67,"tag":225,"props":421,"children":422},{},[423],{"type":73,"value":424},"Pathology and radiology inference",{"type":73,"value":426},": SageMaker async endpoints for\nwhole-slide images and DICOM volumes. Payloads exceed real-time endpoint\nlimits and inference times exceed sync SLOs. Pre-process (tiling,\nresampling) in SageMaker Processing jobs before invoking the endpoint.",{"type":67,"tag":109,"props":428,"children":429},{},[430,435],{"type":67,"tag":225,"props":431,"children":432},{},[433],{"type":73,"value":434},"Genomics pipelines",{"type":73,"value":436},": SageMaker Processing jobs orchestrated by Step\nFunctions for custom workflows; HealthOmics when the workflow fits its\nmanaged WDL\u002FNextflow execution model. Use HealthOmics variant and\nannotation stores for queryable genomic data.",{"type":67,"tag":109,"props":438,"children":439},{},[440,445],{"type":67,"tag":225,"props":441,"children":442},{},[443],{"type":73,"value":444},"Drug discovery and molecular modeling",{"type":73,"value":446},": SageMaker HyperPod for\ndistributed training of large structural or property-prediction models.\nBedrock for molecular description, literature mining, and chemistry Q&A;\nnever for final property prediction where accuracy is load-bearing.",{"type":67,"tag":109,"props":448,"children":449},{},[450,455],{"type":67,"tag":225,"props":451,"children":452},{},[453],{"type":73,"value":454},"Ambient clinical documentation",{"type":73,"value":456},": Bedrock for transcription\npost-processing and structured note generation; guardrails plus\nhuman-in-the-loop sign-off in the EHR.",{"type":67,"tag":76,"props":458,"children":460},{"id":459},"data-architecture",[461],{"type":73,"value":462},"Data Architecture",{"type":67,"tag":83,"props":464,"children":465},{},[466],{"type":73,"value":467},"Pick the store by data shape and access pattern:",{"type":67,"tag":105,"props":469,"children":470},{},[471,481,491,501,511],{"type":67,"tag":109,"props":472,"children":473},{},[474,479],{"type":67,"tag":225,"props":475,"children":476},{},[477],{"type":73,"value":478},"FHIR clinical data",{"type":73,"value":480}," → HealthLake. Supports FHIR-native queries and\nintegrated NLP. Export to S3 for ML training.",{"type":67,"tag":109,"props":482,"children":483},{},[484,489],{"type":67,"tag":225,"props":485,"children":486},{},[487],{"type":73,"value":488},"Observational \u002F claims research",{"type":73,"value":490}," → OMOP CDM on Redshift or Athena.\nStandard vocabulary enables reuse of published phenotype definitions.",{"type":67,"tag":109,"props":492,"children":493},{},[494,499],{"type":67,"tag":225,"props":495,"children":496},{},[497],{"type":73,"value":498},"Genomics",{"type":73,"value":500}," → S3 as the source of truth for VCF, BAM, CRAM, FASTQ;\nLake Formation for governance; Athena for ad-hoc queries; HealthOmics\nstores for indexed access.",{"type":67,"tag":109,"props":502,"children":503},{},[504,509],{"type":67,"tag":225,"props":505,"children":506},{},[507],{"type":73,"value":508},"Medical imaging",{"type":73,"value":510}," → S3 for raw DICOM; HealthImaging for indexed,\nsub-image access and integration with inference pipelines.",{"type":67,"tag":109,"props":512,"children":513},{},[514,519],{"type":67,"tag":225,"props":515,"children":516},{},[517],{"type":73,"value":518},"Unstructured documents",{"type":73,"value":520}," (PDFs, scanned notes) → S3 + Textract for\nextraction, then route structured output to the appropriate store above.",{"type":67,"tag":83,"props":522,"children":523},{},[524],{"type":73,"value":525},"Cross-cutting: Lake Formation for fine-grained access control across\nanalyst personas; Glue Data Catalog as the single metadata plane.",{"type":67,"tag":76,"props":527,"children":529},{"id":528},"security-compliance",[530],{"type":73,"value":531},"Security & Compliance",{"type":67,"tag":83,"props":533,"children":534},{},[535],{"type":73,"value":536},"Treat compliance as a set of hard gates, not a checklist at the end.",{"type":67,"tag":105,"props":538,"children":539},{},[540,550,560,570,580,590],{"type":67,"tag":109,"props":541,"children":542},{},[543,548],{"type":67,"tag":225,"props":544,"children":545},{},[546],{"type":73,"value":547},"HIPAA BAA",{"type":73,"value":549},": Only use services on the AWS HIPAA-eligible list for any\nworkflow touching PHI. Verify eligibility for every service in the\narchitecture, including logging and monitoring sinks.",{"type":67,"tag":109,"props":551,"children":552},{},[553,558],{"type":67,"tag":225,"props":554,"children":555},{},[556],{"type":73,"value":557},"Network isolation",{"type":73,"value":559},": PHI workloads run in private subnets with no\ninternet egress. Use VPC endpoints for S3, SageMaker, Bedrock, KMS, and\nCloudWatch. Enable SageMaker network isolation on training jobs and\nendpoints handling PHI.",{"type":67,"tag":109,"props":561,"children":562},{},[563,568],{"type":67,"tag":225,"props":564,"children":565},{},[566],{"type":73,"value":567},"Encryption",{"type":73,"value":569},": Customer-managed KMS keys for S3, EBS, SageMaker\nvolumes, and Bedrock Knowledge Base vector stores. TLS 1.2 or higher in\ntransit. Scope key policies so only the intended roles can decrypt.",{"type":67,"tag":109,"props":571,"children":572},{},[573,578],{"type":67,"tag":225,"props":574,"children":575},{},[576],{"type":73,"value":577},"Audit and detection",{"type":73,"value":579},": CloudTrail (including data events for PHI\nbuckets), AWS Config rules for drift, GuardDuty for threat detection.\nEnable Bedrock model invocation logging to S3 or CloudWatch — without it\nyou cannot reconstruct what the FM was asked or returned.",{"type":67,"tag":109,"props":581,"children":582},{},[583,588],{"type":67,"tag":225,"props":584,"children":585},{},[586],{"type":73,"value":587},"IAM",{"type":73,"value":589},": Least privilege, no wildcards on PHI buckets or KMS keys.\nSeparate roles for data scientists, training jobs, and inference\nendpoints. Never share a role across environments.",{"type":67,"tag":109,"props":591,"children":592},{},[593,598],{"type":67,"tag":225,"props":594,"children":595},{},[596],{"type":73,"value":597},"De-identification",{"type":73,"value":599},": If the use case permits, de-identify before\ntraining. Document the method (Safe Harbor, Expert Determination) and\nretain the determination evidence.",{"type":67,"tag":83,"props":601,"children":602},{},[603],{"type":73,"value":604},"Compliance decision points to surface explicitly:",{"type":67,"tag":105,"props":606,"children":607},{},[608,613,618],{"type":67,"tag":109,"props":609,"children":610},{},[611],{"type":73,"value":612},"Is any input, output, prompt, or log PHI? If yes, every service in the\npath must be HIPAA-eligible and under BAA.",{"type":67,"tag":109,"props":614,"children":615},{},[616],{"type":73,"value":617},"Will the FM provider see PHI? Confirm the Bedrock model is in an\neligible region and that invocation logs are encrypted with your CMK.",{"type":67,"tag":109,"props":619,"children":620},{},[621],{"type":73,"value":622},"Does the workload cross accounts or regions? Re-verify BAA coverage and\ndata residency commitments.",{"type":67,"tag":76,"props":624,"children":626},{"id":625},"mlops-for-hcls",[627],{"type":73,"value":628},"MLOps for HCLS",{"type":67,"tag":83,"props":630,"children":631},{},[632],{"type":73,"value":633},"Clinical models decay. Assume drift and design for it:",{"type":67,"tag":105,"props":635,"children":636},{},[637,647,657,667,677,687],{"type":67,"tag":109,"props":638,"children":639},{},[640,645],{"type":67,"tag":225,"props":641,"children":642},{},[643],{"type":73,"value":644},"Monitoring",{"type":73,"value":646},": SageMaker Model Monitor for data quality, model quality\n(when labels arrive), and bias drift. Track feature distributions at the\nsite and payer level, not only in aggregate.",{"type":67,"tag":109,"props":648,"children":649},{},[650,655],{"type":67,"tag":225,"props":651,"children":652},{},[653],{"type":73,"value":654},"Distribution shift sources to watch",{"type":73,"value":656},": ICD and CPT coding updates,\nEHR template changes, new site onboarding, seasonal disease patterns,\nformulary changes, care-pathway interventions.",{"type":67,"tag":109,"props":658,"children":659},{},[660,665],{"type":67,"tag":225,"props":661,"children":662},{},[663],{"type":73,"value":664},"Model registry and approvals",{"type":73,"value":666},": Every clinical model goes through the\nSageMaker Model Registry with explicit approval states. Gate promotion\non model card completion, bias evaluation, and clinical sign-off.",{"type":67,"tag":109,"props":668,"children":669},{},[670,675],{"type":67,"tag":225,"props":671,"children":672},{},[673],{"type":73,"value":674},"Experimentation constraints",{"type":73,"value":676},": Randomized online A\u002FB testing on\npatients typically requires IRB approval. Default to shadow mode\nevaluation or champion\u002Fchallenger with human-in-the-loop review.\nDocument the evaluation design before enabling traffic shifts.",{"type":67,"tag":109,"props":678,"children":679},{},[680,685],{"type":67,"tag":225,"props":681,"children":682},{},[683],{"type":73,"value":684},"Reproducibility",{"type":73,"value":686},": Pin container images, training data snapshots\n(S3 versioning), and hyperparameters in the registry. A model that\ncannot be rebuilt cannot be defended to regulators.",{"type":67,"tag":109,"props":688,"children":689},{},[690,695],{"type":67,"tag":225,"props":691,"children":692},{},[693],{"type":73,"value":694},"Incident response",{"type":73,"value":696},": Define rollback criteria and a kill switch for\nevery deployed clinical model before go-live.",{"type":67,"tag":76,"props":698,"children":700},{"id":699},"cost-optimization",[701],{"type":73,"value":702},"Cost Optimization",{"type":67,"tag":83,"props":704,"children":705},{},[706],{"type":73,"value":707},"Optimize only after correctness and compliance are settled:",{"type":67,"tag":105,"props":709,"children":710},{},[711,721,754,764],{"type":67,"tag":109,"props":712,"children":713},{},[714,719],{"type":67,"tag":225,"props":715,"children":716},{},[717],{"type":73,"value":718},"Training",{"type":73,"value":720},": Spot instances for fault-tolerant training; checkpointing\nis mandatory — without it spot interruptions waste compute. HyperPod for\nlong-running distributed jobs to amortize setup.",{"type":67,"tag":109,"props":722,"children":723},{},[724,729,731],{"type":67,"tag":225,"props":725,"children":726},{},[727],{"type":73,"value":728},"Inference shape",{"type":73,"value":730},":\n",{"type":67,"tag":105,"props":732,"children":733},{},[734,739,744,749],{"type":67,"tag":109,"props":735,"children":736},{},[737],{"type":73,"value":738},"Real-time endpoint: latency \u003C ~60s and payload \u003C ~6 MB with steady\ntraffic.",{"type":67,"tag":109,"props":740,"children":741},{},[742],{"type":73,"value":743},"Async endpoint: large payloads (WSI, DICOM volumes, long documents)\nor long inference times.",{"type":67,"tag":109,"props":745,"children":746},{},[747],{"type":73,"value":748},"Serverless inference: bursty or intermittent traffic with cold-start\ntolerance.",{"type":67,"tag":109,"props":750,"children":751},{},[752],{"type":73,"value":753},"Batch transform: offline scoring of large cohorts.",{"type":67,"tag":109,"props":755,"children":756},{},[757,762],{"type":67,"tag":225,"props":758,"children":759},{},[760],{"type":73,"value":761},"Storage",{"type":73,"value":763},": S3 Intelligent-Tiering for research datasets with unknown\naccess patterns; lifecycle policies to Glacier for archival imaging and\ngenomics data retained for compliance.",{"type":67,"tag":109,"props":765,"children":766},{},[767,771],{"type":67,"tag":225,"props":768,"children":769},{},[770],{"type":73,"value":293},{"type":73,"value":772},": Prefer smaller models for extraction and classification;\nreserve larger models for open-ended generation. Cache retrieval results\nwhere clinically safe.",{"type":67,"tag":76,"props":774,"children":776},{"id":775},"decision-trees",[777],{"type":73,"value":778},"Decision Trees",{"type":67,"tag":83,"props":780,"children":781},{},[782],{"type":73,"value":783},"Walk the user through these when the path is not obvious.",{"type":67,"tag":83,"props":785,"children":786},{},[787],{"type":67,"tag":225,"props":788,"children":789},{},[790],{"type":73,"value":791},"Processing clinical text",{"type":67,"tag":140,"props":793,"children":794},{},[795,800,805],{"type":67,"tag":109,"props":796,"children":797},{},[798],{"type":73,"value":799},"Is the task extraction, summarization, or Q&A over text? → Bedrock.",{"type":67,"tag":109,"props":801,"children":802},{},[803],{"type":73,"value":804},"Is it a specialized task (custom NER schema, de-identification at\nscale, domain-specific classifier)? → SageMaker with a fine-tuned\nclinical model.",{"type":67,"tag":109,"props":806,"children":807},{},[808],{"type":73,"value":809},"Is PHI involved? → Apply the Security & Compliance gates regardless of\nwhich service is chosen.",{"type":67,"tag":83,"props":811,"children":812},{},[813],{"type":67,"tag":225,"props":814,"children":815},{},[816],{"type":73,"value":817},"Training on PHI",{"type":67,"tag":140,"props":819,"children":820},{},[821,826,831,836,841],{"type":67,"tag":109,"props":822,"children":823},{},[824],{"type":73,"value":825},"SageMaker training job in a VPC, private subnets only.",{"type":67,"tag":109,"props":827,"children":828},{},[829],{"type":73,"value":830},"Network isolation enabled on the job.",{"type":67,"tag":109,"props":832,"children":833},{},[834],{"type":73,"value":835},"KMS CMK on input S3, training volume, and output artifacts.",{"type":67,"tag":109,"props":837,"children":838},{},[839],{"type":73,"value":840},"BAA-covered region and services end to end.",{"type":67,"tag":109,"props":842,"children":843},{},[844],{"type":73,"value":845},"No internet egress; VPC endpoints for all AWS APIs used.",{"type":67,"tag":83,"props":847,"children":848},{},[849],{"type":67,"tag":225,"props":850,"children":851},{},[852],{"type":73,"value":853},"Choosing an inference pattern",{"type":67,"tag":140,"props":855,"children":856},{},[857,862,867,872],{"type":67,"tag":109,"props":858,"children":859},{},[860],{"type":73,"value":861},"Latency SLO under a few seconds and payload small? → Real-time endpoint.",{"type":67,"tag":109,"props":863,"children":864},{},[865],{"type":73,"value":866},"Payload large (WSI, DICOM, long report) or inference slow? → Async\nendpoint with S3 input and output.",{"type":67,"tag":109,"props":868,"children":869},{},[870],{"type":73,"value":871},"Traffic bursty or low-volume? → Serverless inference.",{"type":67,"tag":109,"props":873,"children":874},{},[875],{"type":73,"value":876},"Scoring a full cohort offline? → Batch transform.",{"type":67,"tag":83,"props":878,"children":879},{},[880],{"type":67,"tag":225,"props":881,"children":882},{},[883],{"type":73,"value":884},"RAG over clinical or scientific content",{"type":67,"tag":140,"props":886,"children":887},{},[888,893,898,903],{"type":67,"tag":109,"props":889,"children":890},{},[891],{"type":73,"value":892},"Is the content stable and well-structured? → Bedrock Knowledge Bases\nwith OpenSearch Serverless.",{"type":67,"tag":109,"props":894,"children":895},{},[896],{"type":73,"value":897},"Chunk by section, not fixed tokens. Preserve tables and dosing blocks.",{"type":67,"tag":109,"props":899,"children":900},{},[901],{"type":73,"value":902},"Require citations in every answer; reject ungrounded responses.",{"type":67,"tag":109,"props":904,"children":905},{},[906],{"type":73,"value":907},"If the corpus includes PHI, ensure the vector store, embedding model,\nand FM are all HIPAA-eligible and under BAA.",{"type":67,"tag":76,"props":909,"children":911},{"id":910},"when-not-to-use-this-skill",[912],{"type":73,"value":913},"When NOT to Use This Skill",{"type":67,"tag":105,"props":915,"children":916},{},[917,922,927],{"type":67,"tag":109,"props":918,"children":919},{},[920],{"type":73,"value":921},"Non-HCLS workloads — this skill's compliance guidance (HIPAA, BAA, FDA SaMD) does not apply to general-purpose applications",{"type":67,"tag":109,"props":923,"children":924},{},[925],{"type":73,"value":926},"Cost estimation or pricing — use the AWS Pricing Calculator; this skill covers architecture patterns, not billing",{"type":67,"tag":109,"props":928,"children":929},{},[930],{"type":73,"value":931},"Reviewing or auditing an existing deployed architecture — this skill designs new systems; use Well-Architected Reviews for existing ones",{"type":67,"tag":76,"props":933,"children":935},{"id":934},"common-mistakes",[936],{"type":73,"value":937},"Common Mistakes",{"type":67,"tag":105,"props":939,"children":940},{},[941,965,986,1007,1028,1049,1070,1091,1112,1133,1154,1175],{"type":67,"tag":109,"props":942,"children":943},{},[944,949,951,956,958,963],{"type":67,"tag":225,"props":945,"children":946},{},[947],{"type":73,"value":948},"Wrong:",{"type":73,"value":950}," Using Bedrock for a problem that needs a custom model\n",{"type":67,"tag":225,"props":952,"children":953},{},[954],{"type":73,"value":955},"Right:",{"type":73,"value":957}," Use SageMaker with a fine-tuned classifier when accuracy is load-bearing and the label space is narrow\n",{"type":67,"tag":225,"props":959,"children":960},{},[961],{"type":73,"value":962},"Why:",{"type":73,"value":964}," FMs underperform well-labeled domain classifiers on constrained tasks, producing unreliable clinical outputs",{"type":67,"tag":109,"props":966,"children":967},{},[968,972,974,978,980,984],{"type":67,"tag":225,"props":969,"children":970},{},[971],{"type":73,"value":948},{"type":73,"value":973}," Using SageMaker when Bedrock would ship in a week\n",{"type":67,"tag":225,"props":975,"children":976},{},[977],{"type":73,"value":955},{"type":73,"value":979}," Start with Bedrock for generic summarization or extraction tasks; move to SageMaker only when forced by accuracy, cost, or data residency\n",{"type":67,"tag":225,"props":981,"children":982},{},[983],{"type":73,"value":962},{"type":73,"value":985}," Building bespoke infrastructure for a commodity NLP task wastes months of engineering time",{"type":67,"tag":109,"props":987,"children":988},{},[989,993,995,999,1001,1005],{"type":67,"tag":225,"props":990,"children":991},{},[992],{"type":73,"value":948},{"type":73,"value":994}," Fixed-size chunking for medical literature in RAG pipelines\n",{"type":67,"tag":225,"props":996,"children":997},{},[998],{"type":73,"value":955},{"type":73,"value":1000}," Chunk by semantic section (abstract, methods, results, guideline recommendation) and preserve tables and dosing blocks intact\n",{"type":67,"tag":225,"props":1002,"children":1003},{},[1004],{"type":73,"value":962},{"type":73,"value":1006}," Fixed chunks split dosing tables, contraindications, and guideline recommendations, destroying retrieval quality",{"type":67,"tag":109,"props":1008,"children":1009},{},[1010,1014,1016,1020,1022,1026],{"type":67,"tag":225,"props":1011,"children":1012},{},[1013],{"type":73,"value":948},{"type":73,"value":1015}," Using real-time endpoints for whole-slide images\n",{"type":67,"tag":225,"props":1017,"children":1018},{},[1019],{"type":73,"value":955},{"type":73,"value":1021}," Use SageMaker async endpoints for WSI and DICOM volumes\n",{"type":67,"tag":225,"props":1023,"children":1024},{},[1025],{"type":73,"value":962},{"type":73,"value":1027}," WSI payloads exceed real-time endpoint size limits and inference times exceed sync timeout SLOs",{"type":67,"tag":109,"props":1029,"children":1030},{},[1031,1035,1037,1041,1043,1047],{"type":67,"tag":225,"props":1032,"children":1033},{},[1034],{"type":73,"value":948},{"type":73,"value":1036}," Forgetting to enable Bedrock invocation logging\n",{"type":67,"tag":225,"props":1038,"children":1039},{},[1040],{"type":73,"value":955},{"type":73,"value":1042}," Always enable Bedrock model invocation logging to S3 or CloudWatch with CMK encryption\n",{"type":67,"tag":225,"props":1044,"children":1045},{},[1046],{"type":73,"value":962},{"type":73,"value":1048}," Without it you cannot reconstruct what the FM was asked or returned — a compliance and debugging failure",{"type":67,"tag":109,"props":1050,"children":1051},{},[1052,1056,1058,1062,1064,1068],{"type":67,"tag":225,"props":1053,"children":1054},{},[1055],{"type":73,"value":948},{"type":73,"value":1057}," Using public subnets or internet egress in PHI training jobs\n",{"type":67,"tag":225,"props":1059,"children":1060},{},[1061],{"type":73,"value":955},{"type":73,"value":1063}," Run PHI workloads in private subnets with no internet egress; use VPC endpoints for all AWS APIs\n",{"type":67,"tag":225,"props":1065,"children":1066},{},[1067],{"type":73,"value":962},{"type":73,"value":1069}," Violates HIPAA isolation expectations even if data itself is encrypted in transit",{"type":67,"tag":109,"props":1071,"children":1072},{},[1073,1077,1079,1083,1085,1089],{"type":67,"tag":225,"props":1074,"children":1075},{},[1076],{"type":73,"value":948},{"type":73,"value":1078}," Using wildcard IAM policies on PHI buckets\n",{"type":67,"tag":225,"props":1080,"children":1081},{},[1082],{"type":73,"value":955},{"type":73,"value":1084}," Scope IAM policies to specific resources and actions; separate roles per environment and persona\n",{"type":67,"tag":225,"props":1086,"children":1087},{},[1088],{"type":73,"value":962},{"type":73,"value":1090}," A single over-scoped role undermines every other security control in the architecture",{"type":67,"tag":109,"props":1092,"children":1093},{},[1094,1098,1100,1104,1106,1110],{"type":67,"tag":225,"props":1095,"children":1096},{},[1097],{"type":73,"value":948},{"type":73,"value":1099}," Training on spot instances without checkpointing\n",{"type":67,"tag":225,"props":1101,"children":1102},{},[1103],{"type":73,"value":955},{"type":73,"value":1105}," Always enable checkpointing when using spot instances for training jobs\n",{"type":67,"tag":225,"props":1107,"children":1108},{},[1109],{"type":73,"value":962},{"type":73,"value":1111}," Spot interruptions lose all progress since last checkpoint; often ends up more expensive than on-demand",{"type":67,"tag":109,"props":1113,"children":1114},{},[1115,1119,1121,1125,1127,1131],{"type":67,"tag":225,"props":1116,"children":1117},{},[1118],{"type":73,"value":948},{"type":73,"value":1120}," Running randomized online A\u002FB tests on patients without IRB approval\n",{"type":67,"tag":225,"props":1122,"children":1123},{},[1124],{"type":73,"value":955},{"type":73,"value":1126}," Default to shadow mode evaluation or champion\u002Fchallenger with human-in-the-loop review\n",{"type":67,"tag":225,"props":1128,"children":1129},{},[1130],{"type":73,"value":962},{"type":73,"value":1132}," Randomized experiments on patients constitute human subjects research — a regulatory violation without IRB",{"type":67,"tag":109,"props":1134,"children":1135},{},[1136,1140,1142,1146,1148,1152],{"type":67,"tag":225,"props":1137,"children":1138},{},[1139],{"type":73,"value":948},{"type":73,"value":1141}," Ignoring distribution shift until aggregate accuracy drops\n",{"type":67,"tag":225,"props":1143,"children":1144},{},[1145],{"type":73,"value":955},{"type":73,"value":1147}," Monitor feature distributions at the site and payer level from day one using SageMaker Model Monitor\n",{"type":67,"tag":225,"props":1149,"children":1150},{},[1151],{"type":73,"value":962},{"type":73,"value":1153}," By the time aggregate metrics move, site-level harm may already have occurred",{"type":67,"tag":109,"props":1155,"children":1156},{},[1157,1161,1163,1167,1169,1173],{"type":67,"tag":225,"props":1158,"children":1159},{},[1160],{"type":73,"value":948},{"type":73,"value":1162}," Assuming de-identification removes all data handling obligations\n",{"type":67,"tag":225,"props":1164,"children":1165},{},[1166],{"type":73,"value":955},{"type":73,"value":1168}," Document the de-identification method (Safe Harbor or Expert Determination) and assess re-identification risk and contractual terms\n",{"type":67,"tag":225,"props":1170,"children":1171},{},[1172],{"type":73,"value":962},{"type":73,"value":1174}," Re-identification risk and contractual obligations often persist after de-identification",{"type":67,"tag":109,"props":1176,"children":1177},{},[1178,1182,1184,1188,1190,1194],{"type":67,"tag":225,"props":1179,"children":1180},{},[1181],{"type":73,"value":948},{"type":73,"value":1183}," Mixing dev and prod in one AWS account with shared KMS keys\n",{"type":67,"tag":225,"props":1185,"children":1186},{},[1187],{"type":73,"value":955},{"type":73,"value":1189}," Separate dev and prod into distinct accounts with independent KMS keys and IAM boundaries\n",{"type":67,"tag":225,"props":1191,"children":1192},{},[1193],{"type":73,"value":962},{"type":73,"value":1195}," Blast radius of any mistake in the shared account covers production PHI",{"type":67,"tag":76,"props":1197,"children":1199},{"id":1198},"when-to-escalate-to-a-human-expert",[1200],{"type":73,"value":1201},"When to Escalate to a Human Expert",{"type":67,"tag":83,"props":1203,"children":1204},{},[1205],{"type":73,"value":1206},"Recommend human review when:",{"type":67,"tag":105,"props":1208,"children":1209},{},[1210,1215,1220,1225],{"type":67,"tag":109,"props":1211,"children":1212},{},[1213],{"type":73,"value":1214},"The use case directly drives a treatment decision without clinician\nreview in the loop.",{"type":67,"tag":109,"props":1216,"children":1217},{},[1218],{"type":73,"value":1219},"The workload requires FDA SaMD classification or 21 CFR Part 11 controls\nyou cannot confirm.",{"type":67,"tag":109,"props":1221,"children":1222},{},[1223],{"type":73,"value":1224},"Data residency, cross-border transfer, or multi-jurisdiction PHI rules\napply.",{"type":67,"tag":109,"props":1226,"children":1227},{},[1228],{"type":73,"value":1229},"A novel FM is being proposed for a safety-critical clinical output with\nno published evaluation on the target population.",{"type":67,"tag":83,"props":1231,"children":1232},{},[1233],{"type":73,"value":1234},"In these cases, draft the architecture, list the open regulatory and\nclinical questions, and stop before implementation.",{"type":67,"tag":76,"props":1236,"children":1238},{"id":1237},"references",[1239],{"type":73,"value":1240},"References",{"type":67,"tag":105,"props":1242,"children":1243},{},[1244,1257,1268,1279],{"type":67,"tag":109,"props":1245,"children":1246},{},[1247,1249],{"type":73,"value":1248},"HIPAA eligible AWS services: ",{"type":67,"tag":1250,"props":1251,"children":1255},"a",{"href":1252,"rel":1253},"https:\u002F\u002Faws.amazon.com\u002Fcompliance\u002Fhipaa-eligible-services-reference\u002F",[1254],"nofollow",[1256],{"type":73,"value":1252},{"type":67,"tag":109,"props":1258,"children":1259},{},[1260,1262],{"type":73,"value":1261},"AWS Well-Architected ML Lens: ",{"type":67,"tag":1250,"props":1263,"children":1266},{"href":1264,"rel":1265},"https:\u002F\u002Fdocs.aws.amazon.com\u002Fwellarchitected\u002Flatest\u002Fmachine-learning-lens\u002Fmachine-learning-lens.html",[1254],[1267],{"type":73,"value":1264},{"type":67,"tag":109,"props":1269,"children":1270},{},[1271,1273],{"type":73,"value":1272},"SageMaker documentation: ",{"type":67,"tag":1250,"props":1274,"children":1277},{"href":1275,"rel":1276},"https:\u002F\u002Fdocs.aws.amazon.com\u002Fsagemaker\u002F",[1254],[1278],{"type":73,"value":1275},{"type":67,"tag":109,"props":1280,"children":1281},{},[1282,1284],{"type":73,"value":1283},"Bedrock documentation: ",{"type":67,"tag":1250,"props":1285,"children":1288},{"href":1286,"rel":1287},"https:\u002F\u002Fdocs.aws.amazon.com\u002Fbedrock\u002F",[1254],[1289],{"type":73,"value":1286},{"items":1291,"total":1467},[1292,1311,1332,1342,1355,1368,1378,1388,1407,1422,1437,1452],{"slug":1293,"name":1293,"fn":1294,"description":1295,"org":1296,"tags":1297,"stars":1308,"repoUrl":1309,"updatedAt":1310},"agentcore-investigation","investigate Bedrock AgentCore runtime sessions","Investigate Bedrock AgentCore runtime sessions via CloudWatch Logs Insights — resolve session\u002Ftrace IDs, query OTEL spans, filter noise, build timelines. Use when debugging AgentCore agent sessions, tracing tool calls, or analyzing latency.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":11},[1298,1299,1302,1305],{"name":27,"slug":28,"type":16},{"name":1300,"slug":1301,"type":16},"Debugging","debugging",{"name":1303,"slug":1304,"type":16},"Logs","logs",{"name":1306,"slug":1307,"type":16},"Observability","observability",9427,"https:\u002F\u002Fgithub.com\u002Fawslabs\u002Fmcp","2026-07-12T08:37:22.601527",{"slug":1312,"name":1313,"fn":1314,"description":1315,"org":1316,"tags":1317,"stars":1308,"repoUrl":1309,"updatedAt":1331},"amazon-aurora-dsql","amazon aurora dsql","build applications with Aurora DSQL","Build with Aurora DSQL — manage schemas, execute queries, handle migrations, diagnose query plans, load data, and develop applications with a serverless, distributed SQL database. Covers IAM auth, multi-tenant patterns, MySQL-to-DSQL and PostgreSQL-to-DSQL schema conversion, FK replacement code generation, OCC retry patterns, ORM migration (Django\u002FHibernate\u002FRails), DDL operations, query plan explainability, SQL compatibility validation, and bulk data loading. Triggers on phrases like: DSQL, Aurora DSQL, create DSQL table, DSQL schema, migrate to DSQL, distributed SQL database, serverless PostgreSQL-compatible database, DSQL query plan, DSQL EXPLAIN ANALYZE, why is my DSQL query slow, DSQL foreign key, DSQL OCC retry, DSQL multi-region, load into DSQL, load CSV into DSQL, bulk load DSQL, aurora-dsql-loader.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":11},[1318,1321,1322,1325,1328],{"name":1319,"slug":1320,"type":16},"Aurora","aurora",{"name":27,"slug":28,"type":16},{"name":1323,"slug":1324,"type":16},"Database","database",{"name":1326,"slug":1327,"type":16},"Serverless","serverless",{"name":1329,"slug":1330,"type":16},"SQL","sql","2026-07-12T08:36:45.053393",{"slug":1333,"name":1334,"fn":1314,"description":1315,"org":1335,"tags":1336,"stars":1308,"repoUrl":1309,"updatedAt":1341},"aurora-dsql","aurora dsql",{"slug":8,"name":9,"logoUrl":10,"githubOrg":11},[1337,1338,1339,1340],{"name":27,"slug":28,"type":16},{"name":1323,"slug":1324,"type":16},{"name":1326,"slug":1327,"type":16},{"name":1329,"slug":1330,"type":16},"2026-07-12T08:36:42.694299",{"slug":1343,"name":1344,"fn":1314,"description":1315,"org":1345,"tags":1346,"stars":1308,"repoUrl":1309,"updatedAt":1354},"aws-dsql","aws dsql",{"slug":8,"name":9,"logoUrl":10,"githubOrg":11},[1347,1348,1349,1352,1353],{"name":27,"slug":28,"type":16},{"name":1323,"slug":1324,"type":16},{"name":1350,"slug":1351,"type":16},"Migration","migration",{"name":1326,"slug":1327,"type":16},{"name":1329,"slug":1330,"type":16},"2026-07-12T08:36:38.584057",{"slug":1356,"name":1357,"fn":1314,"description":1315,"org":1358,"tags":1359,"stars":1308,"repoUrl":1309,"updatedAt":1367},"distributed-postgres","distributed postgres",{"slug":8,"name":9,"logoUrl":10,"githubOrg":11},[1360,1361,1362,1365,1366],{"name":27,"slug":28,"type":16},{"name":1323,"slug":1324,"type":16},{"name":1363,"slug":1364,"type":16},"PostgreSQL","postgresql",{"name":1326,"slug":1327,"type":16},{"name":1329,"slug":1330,"type":16},"2026-07-12T08:36:46.530743",{"slug":1369,"name":1370,"fn":1314,"description":1315,"org":1371,"tags":1372,"stars":1308,"repoUrl":1309,"updatedAt":1377},"distributed-sql","distributed sql",{"slug":8,"name":9,"logoUrl":10,"githubOrg":11},[1373,1374,1375,1376],{"name":27,"slug":28,"type":16},{"name":1323,"slug":1324,"type":16},{"name":1326,"slug":1327,"type":16},{"name":1329,"slug":1330,"type":16},"2026-07-12T08:36:48.104182",{"slug":1379,"name":1379,"fn":1314,"description":1315,"org":1380,"tags":1381,"stars":1308,"repoUrl":1309,"updatedAt":1387},"dsql",{"slug":8,"name":9,"logoUrl":10,"githubOrg":11},[1382,1383,1384,1385,1386],{"name":27,"slug":28,"type":16},{"name":1323,"slug":1324,"type":16},{"name":1350,"slug":1351,"type":16},{"name":1326,"slug":1327,"type":16},{"name":1329,"slug":1330,"type":16},"2026-07-12T08:36:36.374512",{"slug":1389,"name":1389,"fn":1390,"description":1391,"org":1392,"tags":1393,"stars":1404,"repoUrl":1405,"updatedAt":1406},"cost-efficiency-analyzer","analyze cost efficiency and expenses","Analyzes cost structure, cost efficiency, and expense management from P&L data. Use when the user asks about costs, expenses, COGS, operating expenses, cost ratios, cost control, spending efficiency, margin compression from cost side, or wants to understand where money is going. Also use for \"are we spending too much\", \"cost breakdown\", \"expense analysis\", or \"how efficient are our operations\". NOT for revenue or top-line analysis.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":11},[1394,1397,1400,1401],{"name":1395,"slug":1396,"type":16},"Accounting","accounting",{"name":1398,"slug":1399,"type":16},"Analytics","analytics",{"name":702,"slug":699,"type":16},{"name":1402,"slug":1403,"type":16},"Finance","finance",3176,"https:\u002F\u002Fgithub.com\u002Fawslabs\u002Fagentcore-samples","2026-07-12T08:40:03.29555",{"slug":1408,"name":1408,"fn":1409,"description":1410,"org":1411,"tags":1412,"stars":1404,"repoUrl":1405,"updatedAt":1421},"executive-financial-briefing","generate executive financial briefings","Generates a concise executive-level financial briefing or summary suitable for a CEO, CFO, or board presentation. Use when the user asks for a summary, briefing, executive summary, board update, financial overview, financial health check, or \"how is the business doing\". Covers the full P&L picture in one page. Also use for \"give me the highlights\", \"what do I need to know\", or \"quick financial update\".",{"slug":8,"name":9,"logoUrl":10,"githubOrg":11},[1413,1414,1415,1418],{"name":27,"slug":28,"type":16},{"name":1402,"slug":1403,"type":16},{"name":1416,"slug":1417,"type":16},"Management","management",{"name":1419,"slug":1420,"type":16},"Reporting","reporting","2026-07-12T08:40:02.066471",{"slug":1423,"name":1423,"fn":1424,"description":1425,"org":1426,"tags":1427,"stars":1404,"repoUrl":1405,"updatedAt":1436},"multi-quarter-trend-analysis","analyze multi-quarter financial trends","Analyzes financial trends across multiple quarters by comparing P&L metrics over time. Use when the user wants to see trends, patterns, trajectories, or directional movement across 3 or more quarters. Also use for \"how are we trending\", \"show me the trend\", \"track performance over time\", \"quarter over quarter comparison across all quarters\", or any multi-period longitudinal analysis.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":11},[1428,1429,1430,1433],{"name":1398,"slug":1399,"type":16},{"name":1402,"slug":1403,"type":16},{"name":1431,"slug":1432,"type":16},"Financial Statements","financial-statements",{"name":1434,"slug":1435,"type":16},"Variance Analysis","variance-analysis","2026-07-12T08:40:00.79141",{"slug":1438,"name":1438,"fn":1439,"description":1440,"org":1441,"tags":1442,"stars":1404,"repoUrl":1405,"updatedAt":1451},"pdf","process and manipulate PDF documents","Use this skill whenever the user wants to do anything with PDF files. This includes reading or extracting text\u002Ftables from PDFs, combining or merging multiple PDFs into one, splitting PDFs apart, rotating pages, adding watermarks, creating new PDFs, filling PDF forms, encrypting\u002Fdecrypting PDFs, extracting images, and OCR on scanned PDFs to make them searchable. If the user mentions a .pdf file or asks to produce one, use this skill.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":11},[1443,1446,1449],{"name":1444,"slug":1445,"type":16},"Automation","automation",{"name":1447,"slug":1448,"type":16},"Documents","documents",{"name":1450,"slug":1438,"type":16},"PDF","2026-07-12T08:41:44.135656",{"slug":1453,"name":1453,"fn":1454,"description":1455,"org":1456,"tags":1457,"stars":1404,"repoUrl":1405,"updatedAt":1466},"quarterly-kpi-calculator","calculate quarterly financial KPIs","Calculates quarterly financial KPIs from P&L data. P&L figures can be provided directly by the user or fetched from the financial data MCP server. Use when the user wants KPI calculations such as Gross Margin %, EBITDA Margin %, Operating Expense Ratio, or Revenue Growth % QoQ. Also use for quarterly performance review, P&L analysis, or interpreting financial ratios against benchmarks.",{"slug":8,"name":9,"logoUrl":10,"githubOrg":11},[1458,1459,1462,1463],{"name":1395,"slug":1396,"type":16},{"name":1460,"slug":1461,"type":16},"Data Analysis","data-analysis",{"name":1402,"slug":1403,"type":16},{"name":1464,"slug":1465,"type":16},"KPI","kpi","2026-07-12T08:39:59.54971",150,{"items":1469,"total":1556},[1470,1478,1493,1507,1520,1533,1546],{"slug":4,"name":4,"fn":5,"description":6,"org":1471,"tags":1472,"stars":29,"repoUrl":30,"updatedAt":31},{"slug":8,"name":9,"logoUrl":10,"githubOrg":11},[1473,1474,1475,1476,1477],{"name":18,"slug":19,"type":16},{"name":27,"slug":28,"type":16},{"name":21,"slug":22,"type":16},{"name":24,"slug":25,"type":16},{"name":14,"slug":15,"type":16},{"slug":1479,"name":1479,"fn":1480,"description":1481,"org":1482,"tags":1483,"stars":29,"repoUrl":30,"updatedAt":1492},"biomarker-discovery","guide biomarker discovery and validation","Reason about biomarker discovery and validation in HCLS — classifying biomarker intent, choosing feature-selection and cross-validation strategies, avoiding leakage, and planning external replication. Use when the user asks to discover, develop, or validate a biomarker; select features from high-dimensional omics or clinical data; design a validation study; choose evaluation metrics; justify sample size; combine multi-omics signals; or assess clinical utility. Triggers include \"discover a biomarker\", \"validate biomarker\", \"prognostic vs predictive\", \"feature selection\", \"LASSO vs elastic net\", \"nested cross-validation\", \"data leakage\", \"C-index\", \"time-dependent AUC\", \"decision curve analysis\", \"external validation cohort\", \"events per variable\", \"optimism-corrected\", \"multi-omics integration\", \"clinical utility of a biomarker\", \"is this biomarker ready\".",{"slug":8,"name":9,"logoUrl":10,"githubOrg":11},[1484,1485,1488,1489],{"name":27,"slug":28,"type":16},{"name":1486,"slug":1487,"type":16},"Bioinformatics","bioinformatics",{"name":24,"slug":25,"type":16},{"name":1490,"slug":1491,"type":16},"Research","research","2026-07-12T08:37:49.295301",{"slug":1494,"name":1494,"fn":1495,"description":1496,"org":1497,"tags":1498,"stars":29,"repoUrl":30,"updatedAt":1506},"cdisc-compliance","reason about CDISC SDTM and ADaM implementation","Reason about CDISC SDTM and ADaM implementation for regulatory submissions. Use when the user asks about SDTM domain mapping, ADaM dataset design, controlled terminology versioning, define.xml completeness, FDA or PMDA submission requirements, query prioritization by clinical impact, SUPPQUAL usage, or CDISC compliance review. Triggers include \"SDTM mapping\", \"ADaM dataset\", \"CDISC compliance\", \"controlled terminology\", \"define.xml\", \"FDA submission data\", \"PMDA submission\", \"SDTM domain\", \"ADSL\", \"ADAE\", \"ADLB\", \"BDS structure\", \"SUPPQUAL\", \"RELREC\", \"value-level metadata\", \"CDISC CT\", \"regulatory submission data standards\", \"eCTD datasets\", \"SDTM 3.3\", \"ADaM 1.1\", \"query prioritization\", \"clinical data review\".\n",{"slug":8,"name":9,"logoUrl":10,"githubOrg":11},[1499,1502,1503],{"name":1500,"slug":1501,"type":16},"Clinical Trials","clinical-trials",{"name":24,"slug":25,"type":16},{"name":1504,"slug":1505,"type":16},"Regulatory Compliance","regulatory-compliance","2026-07-12T08:37:33.35594",{"slug":1508,"name":1508,"fn":1509,"description":1510,"org":1511,"tags":1512,"stars":29,"repoUrl":30,"updatedAt":1519},"cell-type-annotation","annotate single-cell RNA-seq clusters","Generate code to assign cell type labels to single-cell RNA-seq clusters using CellTypist, SingleR, marker-based annotation, or reference label transfer (scANVI\u002Fingest). Triggers on requests to \"annotate cell types\", \"label clusters\", \"run CellTypist\", \"SingleR annotation\", \"marker gene dotplot\", \"transfer labels from reference atlas\", \"cell identity\", \"automated annotation\", \"reference mapping\", \"scANVI label transfer\", \"canonical markers\", \"immune cell types\", \"hierarchical annotation\", \"majority voting CellTypist\", \"over-clustering annotation\".",{"slug":8,"name":9,"logoUrl":10,"githubOrg":11},[1513,1514,1515,1516],{"name":1486,"slug":1487,"type":16},{"name":1460,"slug":1461,"type":16},{"name":24,"slug":25,"type":16},{"name":1517,"slug":1518,"type":16},"RNA-seq","rna-seq","2026-07-12T08:38:05.443454",{"slug":1521,"name":1521,"fn":1522,"description":1523,"org":1524,"tags":1525,"stars":29,"repoUrl":30,"updatedAt":1532},"cheminformatics","calculate molecular properties with RDKit","Cheminformatics pipeline for small-molecule property calculation, filtering, and similarity analysis using RDKit. Use when the user asks to compute molecular descriptors, filter compounds by Lipinski or Veber rules, detect PAINS, calculate fingerprint similarity, run matched molecular pair analysis, generate ADMET descriptors, or process SMILES. Triggers include \"RDKit\", \"molecular descriptors\", \"Lipinski\", \"rule of five\", \"Veber\", \"PAINS\", \"pan-assay interference\", \"Morgan fingerprint\", \"Tanimoto\", \"fingerprint similarity\", \"matched molecular pair\", \"MMP\", \"mmpdb\", \"ADMET\", \"druglikeness\", \"SMILES\", \"cheminformatics\", \"compound filtering\", \"chemical similarity\".",{"slug":8,"name":9,"logoUrl":10,"githubOrg":11},[1526,1527,1530,1531],{"name":1486,"slug":1487,"type":16},{"name":1528,"slug":1529,"type":16},"Chemistry","chemistry",{"name":1460,"slug":1461,"type":16},{"name":1490,"slug":1491,"type":16},"2026-07-12T08:37:28.334619",{"slug":1534,"name":1534,"fn":1535,"description":1536,"org":1537,"tags":1538,"stars":29,"repoUrl":30,"updatedAt":1545},"claims-analytics","analyze and parse healthcare claims data","Pipeline skill for healthcare claims data parsing, analysis, and fraud detection. Use when the user asks to parse X12 837 or 835 claim files, manipulate ICD-10 CPT or HCPCS codes, detect billing pattern anomalies, profile providers against specialty peers, identify outlier billing behavior, validate NCCI edits programmatically, detect duplicate claims, run Benford's law analysis on charges, build claims data pipelines, or analyze E&M code distributions. Triggers include \"parse X12 837\", \"parse 835\", \"claims SQL\", \"ICD-10 manipulation\", \"CPT code analysis\", \"provider profiling\", \"billing outlier\", \"NCCI validation code\", \"duplicate claim detection\", \"Benford's law charges\", \"claims ETL\", \"E&M distribution analysis\", \"claims analytics pipeline\".\n",{"slug":8,"name":9,"logoUrl":10,"githubOrg":11},[1539,1540,1541,1544],{"name":1460,"slug":1461,"type":16},{"name":21,"slug":22,"type":16},{"name":1542,"slug":1543,"type":16},"Insurance","insurance",{"name":24,"slug":25,"type":16},"2026-07-12T08:37:34.815088",{"slug":1547,"name":1547,"fn":1548,"description":1549,"org":1550,"tags":1551,"stars":29,"repoUrl":30,"updatedAt":1555},"claims-billing-rules","analyze healthcare claims billing rules","Reasoning skill for healthcare claims billing rules and fraud detection logic. Use when the user asks about CMS billing rules, place of service codes, global surgery periods, modifier usage (25 59 76 77), NCCI edit logic, column 1 column 2 code pairs, mutually exclusive procedures, modifier indicators, fraud waste and abuse patterns, E&M upcoding, unbundling, phantom billing, impossible day detection, coding error versus fraud distinction, FWA investigation methodology, or claims audit logic. Triggers include \"CMS billing rules\", \"NCCI edits\", \"modifier 25\", \"modifier 59\", \"global surgery period\", \"upcoding\", \"unbundling\", \"phantom billing\", \"impossible day\", \"FWA\", \"fraud waste abuse\", \"coding error vs fraud\", \"claims audit\", \"billing compliance\", \"E&M level selection\".\n",{"slug":8,"name":9,"logoUrl":10,"githubOrg":11},[1552,1553,1554],{"name":21,"slug":22,"type":16},{"name":1542,"slug":1543,"type":16},{"name":1504,"slug":1505,"type":16},"2026-07-12T08:38:28.210856",40]