Scientific Agent Skills
A field guide to turning a general AI agent into a more disciplined research partner.
Working code isn’t the same as defensible science.
Scientific Agent Skills are small, reviewable procedures that tell an agent how a field expects a task to be done.
An agent can produce plausible code while missing a convention that changes the meaning of the result.
The agent gets explicit guidance about the test, identifier namespace, assumptions and caveats that matter.
“An agent can return working code. A defensible analysis depends on the procedural choices behind it.
Example: RNA-seq
A script may run while treating 100 cells from each of three mice as 300 independent replicates. The skill makes the experimental unit explicit: n = 3 mice, not n = 300 cells.
Example: file coordinates
A pipeline may shift genomic intervals by one base if coordinate conventions are left implicit. A skill turns that convention into a check the agent can follow.
It packages hard-won judgement into something an agent can load when needed.
Faster starts
Skip repeated package and API discovery when a documented workflow already exists.
Better boundaries
Make validation, provenance and safety limits part of the workflow rather than a late reminder.
Reusable practice
Turn a successful research procedure into a portable, inspectable unit that other harnesses can discover.
Think of it as a procedural layer over your existing agent.
The library doesn’t replace your model, tools or expert review. It gives the agent better instructions for using them.
“Use the available skills to query ChEMBL, analyse SAR with RDKit, screen candidates, check resistance mechanisms, and produce a cited report.”
Drug discovery
Compound search, SAR, docking and evidence retrieval.
Omics
Single-cell, RNA-seq, variants and network analysis.
Evidence work
Trials, regulatory documents and traceable literature synthesis.
Methods
Validation, uncertainty, design and reproducibility checks.
Inside the system
What a skill is, how the repository is shaped, and why progressive disclosure is the core design move.
A skill is a directory, not a magical plugin.
At minimum, it has a human-readable SKILL.md. It may also carry references, scripts and assets.
Name, description, licence, compatibility, allowed tools and metadata.
When to use it, what to do, what to check and what not to claim.
Reference material and code stay deferred until the workflow needs them.
Keep the catalogue visible. Load the procedure only when it earns its place.
At the paper’s pinned snapshot, the resident descriptions for all 163 skills used 7.1% of a 200,000-token window.
That is a corpus measurement, not proof that an agent selects correctly.
One research task can compose several kinds of skill.
The library’s worked examples show a pattern: domain skills do the specialised work, while statistics and communication skills connect the workflow into something interpretable.
Find and transform
Retrieve data, resolve identifiers, use a package or query a named scientific resource.
Analyse and check
Apply the appropriate test, assumptions, uncertainty treatment or experimental convention.
Report and trace
Generate figures, explain caveats and return evidence with a provenance trail.
Skills are not a single giant “scientist prompt”. They are composable procedures with different jobs.
The repository has become a broad scientific operating layer.
Current plugin.json version observed in the live repository.
The paper measures the documentation corpus, not agent performance.
Median documented workflow against a 200,000-token window. The second bar exceeds the window.
resident-tier tokens across the 163-skill snapshot.
median documented workflow cost after resident descriptions plus instruction files.
worked workflows that would overflow if every reference file were loaded.
Useful evidence, modest claim.
What the paper supports
- The repository has a tiered documentation design.
- The pinned corpus fits better when references stay deferred.
- Descriptions are a routing bottleneck worth measuring.
- The library is a structured, portable knowledge artefact.
What it does not establish
- That skills improve scientific outcomes on real tasks.
- That a host selects the right skill reliably.
- That every skill is substantively correct or current.
- That portability means identical behaviour across hosts.
Adopt it as a strong starting point, then evaluate the specific skills and workflows that matter to your own research agenda.
The library’s power depends on the agent picking the right procedure.
A long catalogue creates a selection problem before the workflow even starts.
Install a topical subset, not everything by default.
Name the skill when the task is important or ambiguous.
Test selection separately from task completion.
Skills are executable instructions. Treat them like code with influence.
Skills can instruct an agent to run code, install packages, make network requests and modify files. Review what you install.
Community contributions may have uneven review coverage.
Start with skills relevant to the project, not the whole library.
Understand dependencies, tools, credentials and external services.
Record the repository tag, skill name and metadata version used.
Run the recommended scanner and inspect behaviour in your environment.
Make it usable in your harness
A practical path from repository to project-scoped, evaluated workflows across Claude Code, Codex, DeepSeek Harness, Hermes Agent and other hosts.
Start with a narrow research loop.
Choose the work
Pick one repeatable workflow: literature synthesis, methods validation, omics analysis or evidence retrieval.
Install the minimum
Bring in the topical skills your project can actually use. Pin the repo release.
Evaluate the path
Check selection, tool use, provenance, failure modes and the quality of the final artefact.
Use one workflow from your PhD research agenda, such as evidence-traceable literature synthesis, and compare an agent with and without the selected skills.
Don’t start with all 166.
Choose the smallest collection that covers your first workflow. This reduces standing context and the number of routing candidates.
# browse or install the whole repo npx skills add K-Dense-AI/scientific-agent-skills # install one skill directly with GitHub CLI gh skill install K-Dense-AI/scientific-agent-skills scanpy # pin a release for reproducibility gh skill install K-Dense-AI/scientific-agent-skills \ --pin v2.66.0
paper-lookup scientific-writing database-lookup
Traceable retrieval and synthesis before adding domain-specific analysis.
scanpy pydeseq2 scientific-visualization
Pair domain computation with analysis and communication skills.
analytical-method-validation scientific-writing
Make the governing framework and acceptance criteria explicit.
Use the standards path first, then add host-specific wiring.
Claude Code
Use the common skills installer or a project-scoped skills directory. Read the installed SKILL.md.
Codex
For plugin-capable clients, use the repository’s plugin.json plus skills/ layout.
Project scope
Keep the subset beside the project when reproducibility and team review matter.
Verify
Confirm discovery, invoke by name, inspect tool permissions and capture the output provenance.
# Codex plugin route from a local checkout codex plugins install . # Claude Code / Codex: invoke explicitly in a first test “Use the available scientific-agent skill for [workflow]. Name the skill you selected, list its assumptions, and cite the source paths.”
Port the layout, not the assumptions.
If a harness supports the open Agent Skills convention, point it at the same directory. If it doesn’t, build a thin adapter around discovery and invocation.
Place a project-scoped copy under the harness’s documented skills path, or wrap the directory in its native skill/plugin format.
Do not assume Claude/Codex discovery or permissions carry over unchanged.
Use the repository’s documented tap route where supported, otherwise clone into a skills directory the host scans.
Confirm how Hermes handles metadata, credentials, scripts and network access.
Start from the standard directory shape and test one skill end to end before installing a collection.
Portability is intended by the paper, not proven as identical behaviour.
Pair external skills with your own project contract.
Provide reusable scientific procedures and package guidance.
Defines your domain, data handling, naming, evidence and output rules.
Defines credentials, network access, shell permissions and approval gates.
Use this pattern for a research notebook, evidence-traceable literature review, bio-AI convergence scan or reproducibility workflow. Keep personal or clinical data out of a public skills repository.
Ask the agent to expose the procedure, not just the answer.
Use the available scientific skills for this task. Workflow: evidence-traceable literature synthesis Domain: bio-AI convergence Requirements: 1. Name the skills you selected and why. 2. State assumptions and missing data. 3. Keep claims tied to source passages. 4. Separate findings, inference and recommendation. 5. Save a provenance log and an uncertainty note. 6. Stop before clinical or regulatory decisions.
Did the harness find the right skill?
Did the agent follow the procedure?
Did it preserve provenance and caveats?
Evaluate the harness, the skill and the scientific result separately.
Did it choose well?
Create labelled tasks that name the intended skill and close neighbours. Measure correct selection and unnecessary loading.
Did it follow the procedure?
Check tool calls, assumptions, reference loading, script use and stopping behaviour against a rubric.
Did the science hold?
Score against domain answer keys, published conventions, provenance completeness and expert review.
# minimum golden-set record
task_id: literature-001
intended_skill: paper-lookup
must_cite: true
must_state_limits: true
forbidden: patient-specific advice
review: domain-expert + provenance-checkOnce installed, it becomes part of your supply chain.
Repository release, commit or tag. Record each skill’s own metadata version too.
Read the skill and its referenced files before enabling scripts or credentials.
Run your security scanner and inspect changes in pull requests.
Re-run your golden set when the skill or harness changes.
Record the release tag, skill name and metadata.version used. For clinical, regulatory or imaging work, state the specific scope limits.
This makes a changing library identifiable later.
Use Scientific Agent Skills as a procedural starting layer, then earn trust locally.
Start narrow
Select a topical subset around one real workflow.
Make it visible
Ask the agent to name its skill, assumptions, sources and limits.
Measure what matters
Evaluate selection, adherence, provenance and scientific outcomes separately.
Skills don’t turn an agent into a scientist by themselves. They give it a better procedural memory, which you still need to test, constrain and review.
Read the source, pin the version, keep the claim modest.
Scientific Agent Skills repository
Live README, plugin metadata, installation guidance, examples and security disclaimer.
Scientific Agent Skills: A Library of Procedural Knowledge for Research Agents
Kassis et al., arXiv:2609.00065, v2, September 2026.
Agent Skills standard
The open convention the repository targets.
- The paper measures the documentation corpus, not task outcomes.
- The live repository moves faster than the paper’s pinned snapshot.
- Host support and discovery behaviour need local verification.
- Scientific and clinical outputs still need qualified review.
Deck prepared from publicly available repository and paper material. Implementation commands are reproduced as guidance from the source and should be checked against the current host version before use.