Life Science Research

Life-Science Research Router

This is a portable routing skill for Claude Code, Codex, Gemini CLI, and other agents. It distills the useful pattern from Codex's Life Science Research plugin: a small router plus many source-specific lookup recipes. Do not treat it as a single monolithic "biomed answer engine."

Concrete source-specific recipes are nested under sources/ inside this skill. Each is a directory with an INSTRUCTIONS.md (renamed from SKILL.md so harness skill-discovery does not surface all 50 as top-level skills — that blows the skill-description budget in Claude Code and Codex). When a narrow source lookup matches the request, open only that source's sources/<source-skill>/INSTRUCTIONS.md and run its local helper script.

Lessons From Codex's Life Science Plugin

Copy the architecture, not the package.

What is worth copying:

• Router first: classify the question into 1-3 evidence lanes before searching.
• Normalize entities before evidence: gene/protein, variant, disease, compound, pathway, and dataset identifiers decide which source is valid.
• Thin source recipes: each database wrapper is small, explicit, and biased toward compact JSON summaries rather than broad dumps.
• Small windows by default: start with 5-10 records, then expand only after seeing signal.
• Raw payload escape hatch: save large raw responses to a file/path when needed; do not paste source dumps into the agent context.
• Freshness discipline: re-run important external lookups in long sessions instead of trusting stale context.
• Output contract: synthesize around the user's question, then separate database facts, literature claims, inference, caveats, and next checks.

What not to copy:

• 50 source wrappers as a default local dependency.
• Plugin-specific tool names in portable instructions.
• A monolithic answer engine that hides which sources were checked.
• Advisory behavior buried behind an MCP call that agents may forget to make.

Agent-Infra Placement Rule

Use the lowest durable layer that solves the problem:

For this skill, the right shape is: one visible shared router plus nested source wrappers. Do not expose every source wrapper as a top-level skill unless a specific agent cannot reliably discover the router and the benefit is measured.

Quick Route (start here)

Before anything else, decide which entry point fits the question shape. This table compresses the lane + tool tables below into a single decision.

Do NOT scan the full Source Index until the route above falls through.

Biomedical MCP coverage hint

If the host agent has the biomedical MCP loaded, the following sources are already wrapped — prefer the MCP tool over opening the source skill: HPO / OMIM / Orphanet / Monarch (phenotype_*, rare_disease_*), PanelApp (panels_*), ClinGen gene validity + dosage (curation_*), KEGG (pathways_kegg_*), CPIC (folded into supplements_pharmgkb_lookup / targets_pharmacogenetics), DGIdb (via targets_*), LitVar2 (literature_variant_publications), ISBT blood groups (bloodgroups_*), GTEx eQTL (expression_eqtl), USDA FoodData (nutrition_*). Open the per-source skill only if (a) the MCP isn't loaded, (b) the MCP wrapper is too coarse for the question, or (c) you need a query the wrapper doesn't expose.

Core Rule

Start by deciding the evidence lane, normalize entities, then call the smallest set of authoritative sources that can answer the question. Synthesize the answer around the user's question, not around the tools.

For personal health/genomics projects, local curated state wins over fresh web search:

1. repo-owned entity pages, clinical maps, claim stores, and WGS artifacts
2. typed MCP / CLI views over those artifacts
3. external biomedical databases
4. literature and web search

Old AI-chat recall is not evidence unless confirmed by a stronger source.

Routing Lanes

Pick 1-3 lanes. Expanding beyond that is usually a landscape review.

Entity Normalization

Normalize before deep retrieval.

If identifiers disagree, stop and resolve the conflict before interpretation.

Tool Preference By Agent Surface

Use whichever surface exists in the current agent. Do not assume all tools are available everywhere.

When no MCP is exposed, use local CLIs or direct APIs. State the fallback in the answer if it affects confidence.

Source Index

Open only the sources/<name>/INSTRUCTIONS.md you need. Each source has a local helper script under sources/<name>/scripts/.

Endpoints without a local wrapper yet

These have public APIs / downloads but no sources/<name>/INSTRUCTIONS.md recipe in this skill yet. Use the endpoint directly until a wrapper lands.

Retrieval Discipline (per-source)

Request-time cautions that matter more than source identity:

• ClinVar: separate star/review status from clinical significance.
• gnomAD: report population AF with build + ref/alt before any pathogenicity claim.
• Open Targets: distinguish genetics evidence from literature/mining scores.
• GWAS Catalog: carry ancestry and study context into the answer.
• GTEx / eQTL Catalogue: report tissue and variant build.
• Human Protein Atlas / Bgee: tissue expression is not causality.
• Reactome / QuickGO: verify species and evidence code where available.
• STRING: network topology = interaction context, not proof of mechanism.
• AlphaFold / RCSB PDB: check confidence; do not overstate low-confidence regions.
• ChEMBL / BindingDB: report assay type, organism, units, relation.
• PharmGKB / CPIC: attach evidence/guideline level.
• ClinicalTrials.gov: trial existence ≠ efficacy evidence.
• cBioPortal / CIViC: carry tumor type and evidence level.
• PubMed / PMC / bioRxiv: preprints are provisional. Always check retraction / correction status for any paper that drives a clinical or mechanistic claim — query api.crossref.org/v1/works/{DOI} and inspect update-to[] for type: retraction or type: correction. Retrieval can otherwise launder retracted sources into confident answers.
• Guideline currency: when citing CPIC / ACMG / professional-society guidance, name the version and year — older guideline versions are routinely superseded (e.g., ACMG 2015 → 2023, CPIC v1 → v2 per gene). Don't anchor on a guideline number without checking PharmGKB/CPIC for the current revision.

Execution Discipline

1. Write the retrieval plan in one short paragraph if the task is broad.
2. Prefer direct lookups before expensive multi-step chains.
3. Use small result windows first: 5-10 records, then expand only if needed.
4. Save or cite raw payload paths when results are large; do not paste dumps.
5. Re-run important external lookups in long sessions instead of trusting old context.
6. Separate database facts, literature findings, inference, and advice.
7. Attach caveats for ancestry, cohort, tissue, assay, model organism, and study design.
8. Never turn mechanistic plausibility into treatment advice.

Parallelization

Parallelize only independent lanes. Keep entity normalization, scope decisions, conflict resolution, and final synthesis with the coordinating agent.

Good parallel splits:

• genetics vs expression vs structure vs pharmacology for one gene
• multiple variants using the same output schema
• literature review vs database annotation
• clinical trials vs molecular mechanism

Bad splits:

• one narrow lookup
• tasks where later queries depend on unresolved identifiers
• broad fan-out before deciding what entity is being studied

Each worker should return:

• what it checked
• key findings
• caveats
• tools/sources used
• whether the result is database fact, literature claim, or inference

Output Contract

Default answer shape:

1. direct answer or working conclusion
2. evidence by lane
3. conflicts or missing evidence
4. next useful lookup or validation step

For personal health/genomics answers, add:

• what local artifact was used
• whether external evidence updates or weakens the local state
• whether the conclusion is action-grade, research-only, or null

For any literature-anchored claim, also surface:

• publication year + retraction/correction status (CrossRef update-to[])
• guideline version where applicable (ACMG / CPIC / professional society)
• whether the cited evidence is on current frontier models when the claim is about LLM behavior — pre-frontier findings (GPT-3.5/4, Claude 3, Gemini 1.x) don't transfer unless the result is scale-independent.

When To Stop

Stop early when:

• identifier normalization fails
• local curated docs already contain a newer reviewed answer
• external evidence is only associative and the user's question asks for causality
• the remaining sources are lower tier than what has already been checked
• the likely next step is a repo-owned contract update, not more searching