Claude uses wrong function name from a skill's library

Tell it to re-read the skill's reference folder; the SKILL.md frontmatter should auto-load it.

scientific-agent-skills — インストール & ライブデモ

なぜ使うのか

主な機能

133 skills spanning genomics, cheminformatics, molecular dynamics, clinical trials, ML training, and paper writing
REST API access recipes for 78+ public databases (ClinVar, COSMIC, ClinicalTrials.gov, FDA, PubMed, bioRxiv, arXiv)
Each skill includes runnable code examples and links to primary documentation
Claude routes between skills automatically based on the task domain

ライブデモ

実際の動作

scientific-agent-skill.replay ▶ 準備完了

0/0

インストール

作者のリポジトリ · Claude Skills の公式ドキュメント

ユースケース

実用的な使い方： scientific-agent-skills

Run a single-cell RNA-seq analysis with Scanpy

👤 Computational biologists and bioinformatics postdocs ⏱ ~30 min intermediate

使うタイミング： You have an .h5ad file and want QC, clustering, and UMAP without re-reading the Scanpy docs.

前提条件

Scanpy installed in your env — pip install scanpy
An .h5ad dataset — Download from a public repo or use your own

フロー

Tell Claude the data and the goal

I have pbmc3k.h5ad. Use the scanpy skill to do QC, normalization, clustering, and a UMAP. Explain each step.✓ コピーしました

→ Claude calls the right Scanpy functions in the right order with sensible defaults
Iterate on parameters

Re-do clustering with resolution 0.8 and show how cluster labels changed.✓ コピーしました

→ Parameter tweak without re-planning the full pipeline

結果： A reproducible notebook plus figures, using idiomatic Scanpy — no hallucinated function names.

注意点

Generic ML advice from Claude that ignores domain conventions — Reference the skill by name: 'use the scanpy skill'

組み合わせ： filesystem

Build a structured literature review from PubMed + arXiv

👤 PhD students, research scientists ⏱ ~45 min intermediate

使うタイミング： You need 30+ papers on a topic with abstracts and citation metadata, not just titles from Google.

フロー

Specify scope

Use the paper-lookup skill. Find all PubMed and bioRxiv papers on GLP-1 agonists for Alzheimer's in the last 3 years.✓ コピーしました

→ Claude hits the right APIs with correct query syntax
Cluster by theme

Cluster by hypothesis (neuroinflammation, vascular, direct neuronal) and give me the top 3 papers per cluster.✓ コピーしました

→ Thematic grouping with citation-ready metadata

結果： A reviewable literature map with real DOIs, not fabricated citations.

注意点

Claude hallucinates paper titles without the skill — Always verify DOIs — have Claude fetch one record to prove it's real

組み合わせ： fetch

Dock a small molecule to a target protein with DiffDock

👤 Medicinal chemists, drug discovery researchers ⏱ ~20 min advanced

使うタイミング： You have a SMILES string and a PDB target and want a quick first-pass pose prediction.

前提条件

DiffDock environment — Follow the DiffDock skill's env-setup recipe

フロー

Provide ligand and receptor

Using the diffdock skill, dock SMILES 'CC(=O)OC1=CC=CC=C1C(=O)O' into PDB 1ABC. Give me the top 5 poses with scores.✓ コピーしました

→ Claude runs the right DiffDock command with correct flags
Visualize

Generate a PyMOL script to render the top pose.✓ コピーしました

→ Runnable .pml file

結果： Pose predictions ready to feed into a downstream free-energy calculation.

注意点

Claude invents flags that don't exist in DiffDock — The skill's reference folder has the real CLI — insist Claude consult it

組み合わせ： filesystem

組み合わせ

他のMCPと組み合わせて10倍の力を

scientific-agent-skill + filesystem

Skills produce notebooks and figures; filesystem MCP stores and organizes the output

Save the Scanpy UMAP to results/figures/ and write a README for the experiment.✓ コピーしました

scientific-agent-skill + arxiv

Pair the paper-lookup skill with the arxiv MCP for richer citation graphs

Find arXiv papers citing the one I just read, grouped by which section they cite.✓ コピーしました

ツール

このMCPが提供する機能

ツール	入力	呼び出すタイミング	コスト
Bioinformatics (BioPython, Scanpy, pysam, gget, scVelo)	sequence / count matrix / BAM file	Sequence, genomics, or single-cell workflow	0 — local compute
Cheminformatics (RDKit, Datamol, DeepChem, DiffDock, OpenMM)	SMILES / PDB / MOL2	Small molecule, protein structure, or dynamics problem	0
Clinical databases (ClinVar, COSMIC, ClinicalTrials.gov, FDA)	gene / variant / trial ID	Looking up clinical evidence or trial status	0 — public APIs
Paper lookup (PubMed, bioRxiv, arXiv)	query string, date range	Literature search with real citations	0
ML training (PyTorch Lightning, Transformers, PyMC, TimesFM)	dataset + config	Building a model with idiomatic framework usage	0

コストと制限

運用コスト

APIクォータ: Public databases (NCBI, PubMed) have their own polite-use limits — typically 3 req/s
呼び出しあたりのトークン: SKILL.md references are sizeable; expect 2-5k tokens loaded per domain touched
金額: Free — skills are local files; only pay for compute you run
ヒント: Scope the prompt to one domain at a time so Claude doesn't load every SKILL.md.

セキュリティ

権限、シークレット、影響範囲

認証情報の保管： No credentials in the skill itself; NCBI/PubMed work better with an email-in-env-var for polite rate limiting

データ送信先： Only to the public science APIs you choose to query

Clinical data skills can hit PHI-adjacent records — make sure your usage complies with IRB/HIPAA when applicable.

トラブルシューティング

よくあるエラーと対処法

API rate limit hit on NCBI

Add an email to NCBI_EMAIL env var and cap concurrency at 3 req/s.

Skill not invoked on a relevant prompt

Mention the library name explicitly ('use the rdkit skill') — auto-invocation is fuzzy across 133 skills.

代替案

scientific-agent-skills 他との比較

代替案	代わりに使う場面	トレードオフ
ai-research-skill (Orchestra-Research)	You want ML-research infrastructure (training, distributed, inference) rather than wet-lab / clinical	More ML systems, less biology / chemistry domain depth
biomcp	You want an MCP server (live tools) rather than prompt-bundle skills	MCPs execute calls; skills just teach Claude to write the right code

その他

リソース

📖 GitHub の公式 README を読む

🐙 オープンな issue を見る

🔍 400以上のMCPサーバーとSkillsを見る