---

name: medical-research-skills description: "Route medical and biomedical research requests across the aipoch medical-research-skills catalog, its evidence, protocol, analysis, writing, operations, privacy, and audit workflows; use the bundled offline index and preserve explicit safety and verification boundaries." disable-model-invocation: true metadata: disco-role: operating license: MIT

Medical Research Skills

Use this repo skill as a catalog router, not as a replacement for every specialist skill in the collection. It helps a Researcher select a bounded medical-research workflow, identify the right category and collection, inspect available entry points, and state what still needs specialist dependencies, source evidence, credentials, human review, or backend verification.

When to use

Read this route when a request:

• asks which medical-research skill to use, or names a topic, modality, database, analysis, manuscript artifact, or clinical-research workflow;
• spans multiple stages such as evidence → protocol → analysis → writing;
• needs an offline catalog search by task phrase, canonical skill id, category, API, data format, or model signal;
• asks to audit, install, export, or troubleshoot a skill from this repository.

Do not treat a catalog match as proof that a medical claim, statistical result, clinical recommendation, citation, API result, or generated code is correct. For diagnosis, prescribing, patient-specific decisions, or regulated use, stop at research support and require qualified human review.

Route by task shape

1. Evidence discovery and appraisal → sub-skills/evidence-insight/SKILL.md for search design, retrieval, paper reading, evidence maps, gaps, and citation provenance.
2. Study or experiment planning → sub-skills/protocol-design/SKILL.md for aims, design, eligibility, endpoints, bias, power, validation, and feasibility.
3. Data and model execution → sub-skills/data-analysis/SKILL.md for schemas, clinical statistics, omics, imaging, survival, diagnostic ML, visualization, and reproducibility.
4. Manuscript and submission artifacts → sub-skills/academic-writing/SKILL.md for sections, reviews, abstracts, figures/tables, citations, reporting, peer-review, grants, and journal adaptation.
5. Utilities, privacy, audit, installation, and export → sub-skills/operations-and-audit/SKILL.md for PHI/privacy, documents, lab operations, deterministic utilities, quality gates, and side-effectful installation decisions.

For a multi-stage request, keep ownership explicit: evidence defines what is known, protocol defines what will be studied, analysis defines how data are processed, and writing communicates verified results. Link the handoffs rather than blending unsupported claims across stages.

Offline catalog search

The bundled references/catalog-index.json is the canonical discovery index for this generated graph. It records 604 entries from the two public collections, their five categories, descriptions, task signals, and whether the source entry advertised references, scripts, requirements, or assets. Read the index for selection; do not infer runtime support from a name alone.

Run the deterministic, network-free helper when a query is broad:

python scripts/catalog_query.py "bulk RNA-seq differential expression" --limit 5
python scripts/catalog_query.py "PubMed search" --category "Evidence Insight" --json
python scripts/catalog_query.py --collection scientific-skills --limit 10

The helper returns candidate ids and descriptions only. A zero-result exit is a routing signal: broaden the phrase, remove an over-specific filter, or ask for the missing task/data/backend details. It does not execute a matched skill. Read references/catalog-usage.md for collection selection, confidence, and handoff fields.

Safe operating sequence

1. Normalize the requested artifact, input data/source, target audience, environment, time/compute/network limits, and acceptance criteria.
2. Query the index and shortlist one primary route plus, only when needed, one supporting route. Prefer the narrowest route whose inputs and outputs match.
3. Record assumptions, evidence/provenance requirements, optional dependencies, credentials, backend needs, PHI exposure, and stop conditions before running anything.
4. Read the selected sub-skill reference, validate the input schema, and choose an instruction-only or script-backed specialist entry. Never run arbitrary user strings through a shell, eval, or exec.
5. Separate deterministic checks from model/domain judgment. Preserve raw input, generated code, intermediate artifacts, and validation results where the user needs reproducibility.
6. Return deliverable, assumptions, evidence used, limitations, unresolved checks, and the next verification step. Do not fabricate papers, identifiers, statistics, patient facts, API responses, or completed execution.

Installation baseline and side effects

This repository is a heterogeneous Markdown/Python/R/Node catalog, not one installable Python distribution: do not run pip install . at the catalog root. The generated graph deliberately ships only standard-library discovery and check helpers, so the minimal catalog-level verification is the Python compile/help sequence below. Individual catalog entries may need separate Python packages, R packages, Node tools, network access, API keys, clinical datasets, model weights, or CUDA/other accelerators. Prepare only the minimum environment for the selected specialist workflow; do not install every per-entry requirements file.

For a selected specialist, read its own entry and requirements first, create an isolated environment, install only the documented dependency variant, and run its smallest parser/import/tiny-fixture check before any expensive workflow.

Before using any installer, external API, data download, or directory mutation, read references/troubleshooting.md and the operations route. Installation/export is never implicit. The original OpenClaw installer is documented as a reference contract, not bundled or run by this skill.

Verification

Run the offline graph check after changing catalog content, routes, or helper scripts:

python scripts/check_catalog_skill.py
python -m py_compile scripts/catalog_query.py scripts/check_catalog_skill.py
python scripts/catalog_query.py "survival analysis" --limit 3

For final production verification, use the configured review artifact directory and verify-repo-skill. Native/API/GPU/R/Node/credentialed/expensive workflows remain optional candidates and must be marked unverified when not run. Read references/repo-provenance.md before deciding whether a later checkout requires refresh.