菜单
Deterministic ClinVar classification lookup for the patient's variants. Pure code, no LLM. Emits Tier A for Pathogenic/Likely-Pathogenic at >=2-star review, Tier B for VUS.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
Open Targets gene-disease association lookup via the official MCP connector. One GraphQL round trip per gene. Emits Tier C informational evidence.
Standing watcher. Reruns the trial, literature (PubMed + bioRxiv), and ClinVar-relookup specialists against the FIXED profile and diffs against the stored baseline. Empty diff -> one calm "no significant changes" timeline entry.
PubMed + bioRxiv search for recent literature on the patient's genes and diagnosis. bioRxiv preprints are first-class (earliest signal). Emits Tier C; the LLM only summarizes genuine free text, it never asserts facts.
ClinicalTrials.gov v2 search for trials relevant to the patient's diagnosis and genes. Eligibility is surfaced as "may be eligible — confirm with your care team", never confirmed. Emits Tier C.
Uncertainty-ranked gap report (spec §8b). For an unresolved (Tier B) variant, names that the analysis can't resolve it and attributes the gap to a missing input — a documented family history — that would let the analysis weight it. Never a test order.
Uncertainty-ranked gap report (spec §8b). Names a pharmacogenomic module the analysis could NOT complete and attributes the gap to a specific missing input (a genotype). Never tells the patient to get a test — it is a report on the tool's blind spots for the care team.
| name | variant-clinvar-lookup |
| kind | oneshot |
| domain | variant |
| description | Deterministic ClinVar classification lookup for the patient's variants. Pure code, no LLM. Emits Tier A for Pathogenic/Likely-Pathogenic at >=2-star review, Tier B for VUS. |
A deterministic specialist (spec §7). It does not call an LLM. For each variant
on the profile it looks up the ClinVar classification and review status and emits an
EvidenceRecord with a fixed tier:
The classification comes from app.clinvar.lookup, which reads the local ClinVar
source (demo JSON now, tabix-indexed VCF in production). No fact here is asserted by
a model, so none can be hallucinated.
Entrypoint: specialist.run_specialist(profile, ctx).
