Science Skills
This repository contains the prompt, skill, agent, and MCP server assets for a
Claude Science style scientific computing assistant. The contents are mostly
runtime-facing configuration and documentation: skills teach the assistant how
to perform domain workflows, agents define specialized profiles, and MCP
servers expose scientific data and chemistry tools.
Source
The base prompt and bundled skills in this repository are sourced from the
Claude Science product. The public product page is available for context:
https://claude.com/product/claude-science.
Repository Layout
| Path | Purpose |
|---|
SYSTEM_PROMPT.md | Base system prompt and operating rules for the scientific computing agent. |
skills/ | Skill packages. Each skill is centered on a SKILL.md file, with optional helper code, references, requirements, and provider metadata. |
agents/ | Agent profile metadata and prompts for specialized roles such as onboarding, transcript review, bookmarking, and the general science agent. |
mcp-servers/ | Bundled MCP servers and widgets used by the agent for external scientific data and chemistry workflows. |
Skill Catalog
The skills/ directory includes workflows for several broad areas:
- Biomolecular modeling and design:
alphafold2, boltz, chai1,
openfold3, esmfold2, fair-esm2, diffdock, proteinmpnn,
ligandmpnn, solublempnn, evo2, and borzoi.
- Scientific analysis and reporting:
literature-review, pdf-explore,
figure-style, figure-composer, paper-narrative, and
indication-dossier.
- Single-cell and omics workflows:
scgpt and scvi-tools.
- Runtime and environment operations:
remote-compute-modal,
remote-compute-ssh, compute-env-setup, managed-model-endpoints, and
using-model-endpoint.
- Product and customization workflows:
customize, skill-creator,
product-self-knowledge, and self-awareness.
Most skills start with YAML front matter:
---
name: example-skill
description: When and why this skill should be loaded.
license: Apache-2.0
---
The body of SKILL.md should give the operational procedure, expected inputs,
failure modes, validation steps, and any tool-specific constraints needed for
reliable execution.
Agents
Agent profiles live under agents/<name>/metadata.yaml.
operon is the general-purpose scientific computing agent.
onboarding supports first-run user onboarding.
reviewer audits another agent transcript for unsupported claims,
fabrication, or plan deviation.
bookmarker selects transcript spans worth preserving as navigation
breadcrumbs.
These files are configuration assets, not standalone applications. Keep prompt
changes tightly scoped and preserve any eval or benchmark notes near the prompt
they justify.
MCP Servers
The mcp-servers/ directory contains bundled MCP integrations:
bio-tools vendors multiple Python packages for biological and biomedical
data retrieval. run_server.py launches a named stdio server from
mcp-servers/bio-tools/lib/.
ketcher-chemistry contains a chemistry MCP server and widget assets for
structure editing and chemistry workflows.
The MCP code is runtime infrastructure. When changing it, prefer narrow edits
and verify both the server entry point and the shape of returned tool data.
Working With This Repo
Useful inspection commands:
# List all skills
find skills -maxdepth 2 -name SKILL.md | sort
# List agent profiles
find agents -maxdepth 2 -name metadata.yaml | sort
# Show launchable bio-tools MCP servers
find mcp-servers/bio-tools/lib -maxdepth 2 -name server.py \
| sed 's#^mcp-servers/bio-tools/lib/##; s#/server.py$##' \
| sort
Before editing a skill or prompt:
- Read the relevant
SKILL.md or metadata.yaml completely.
- Preserve existing safety, artifact, provenance, and validation rules unless
the change explicitly updates them.
- Keep examples executable and aligned with the current runtime surface.
- Add or update references next to the skill when a workflow depends on
external documentation or non-obvious operational knowledge.
Maintenance Notes
- Use ASCII in new files unless the surrounding file already requires a wider
character set.
- Avoid broad rewrites of prompts or generated MCP bundles without a concrete
regression target.
- Treat bundled third-party services, model weights, and APIs as dependencies:
document licenses, terms, privacy implications, and data movement in the
relevant skill metadata when applicable.
- If a skill ships helper code such as
kernel.py, keep the markdown
instructions and helper APIs in sync.