Scientific Agent Skills

Scientific Agent Skills

Use this skill as the safe discovery, provenance, routing, and selective-install front door for K-Dense-AI/scientific-agent-skills. The upstream repository is a large collection of independent scientific skills, not one scientific runtime. Do not load, copy, or install every folder merely because the repository was named.

The audited snapshot is release v2.65.0, commit f6fcafeb1cc8c82eca0160a18bc41c38427b8e0f. It contains 163 directories with a SKILL.md. Treat that count as a pinned observation, not a permanent fact. Re-audit a newer tag or commit before quoting its inventory.

When to use this skill

• Inspect, pin, route, selectively install, refresh, or troubleshoot the named K-Dense scientific skill collection.
• Choose one narrow upstream owner for a scientific package, database, method, lab platform, analysis workflow, or communication artifact.
• Review frontmatter, support files, executable helpers, dependencies, credentials, network destinations, licenses, and destination collisions before installation.
• Compare a local selection with a newer explicit upstream tag or commit.
• Separate instruction-file installation from the scientific, cloud, clinical, publication, or hardware action that an installed skill later proposes.

Do not use this wrapper for nearby jobs that already have a local owner:

• End-to-end scholarly discovery, writing, review, and publication: academic-research.
• General web investigation and structured evidence collection: deep-research.
• Drafting or revising a research paper: research-paper-writing.
• Academic diagrams and publication figures: paperbanana.
• Scientific reasoning benchmark selection: scientific-llm-benchmarks.
• Existing local Word, PDF, PowerPoint, or Excel work: the local docx, pdf, pptx, or xlsx skill. Never replace those with the restricted upstream copies.

Instructions

Step 1: Pick one operating mode

Do not blend install with operate. Installing instructions does not approve package installation, data upload, an API call, a cloud job, laboratory control, or a clinical or publication output.

Step 2: Establish provenance and the license boundary

1. Prefer an existing trusted checkout. Otherwise clone to a staging directory.
2. Record the exact commit and origin. Never treat moving main as a pin.
3. Verify the root MIT notice and inspect every selected skill's own license field and bundled license file.
4. Run the bundled helper without executing any upstream skill script:

python3 .agent-skills/scientific-agent-skills/scripts/audit-pack.py doctor \
  --repo /path/to/scientific-agent-skills \
  --expect-commit f6fcafeb1cc8c82eca0160a18bc41c38427b8e0f \
  --format json

The helper reads files and Git metadata only. It reports tree, README, and docs/skills.md inventory counts, frontmatter, support-code volume, declared licenses, symlinks, and coarse risk signals. A risk hit is a review lead, not proof that a skill is malicious or safe. At the audited pin, WARN is expected: the real tree and README report 163 skills, while docs/skills.md lists 162 and omits waypoint-bio.

Four folders are a hard redistribution boundary at the audited pin: docx, pdf, pptx, and xlsx. Their bundled Anthropic terms prohibit retaining, copying, deriving, and redistributing the materials outside the allowed services. Do not copy, adapt, install through this wrapper, or publish them in jeo-skills. Route those formats to the existing local skills.

pacsomatic has a separate MIT notice from Beifang Niu. Other skills can also declare licenses that differ from the repository root. Preserve the selected folder's actual attribution and terms.

Read source audit and risk before any copy or install decision.

Step 3: Route to one narrow upstream skill

Use catalog and routing as the pinned inventory. Common lanes include:

Select by the user's actual input, output, scientific domain, and execution surface. Do not load several overlapping skills as a substitute for deciding. If the request is a general research pipeline rather than operation of the K-Dense pack, use the local route-out instead.

Step 4: Preview a selective installation

Use a detached or clean checkout at the reviewed pin:

git clone --filter=blob:none \
  https://github.com/K-Dense-AI/scientific-agent-skills.git \
  /path/to/scientific-agent-skills
git -C /path/to/scientific-agent-skills checkout \
  f6fcafeb1cc8c82eca0160a18bc41c38427b8e0f

# Read-only inventory from the reviewed checkout
npx --yes skills@1.5.23 add /path/to/scientific-agent-skills \
  --list --full-depth

# Read-only destination and license/collision plan
python3 .agent-skills/scientific-agent-skills/scripts/audit-pack.py plan \
  --repo /path/to/scientific-agent-skills \
  --target /path/to/agent/skills \
  --skill scanpy \
  --format json

The plan never creates the target. It blocks unknown names, invalid frontmatter, existing destinations, and the four restricted document folders. Do not use --all as a shortcut: it crosses license boundaries, adds excessive standing context, and hides dependency conflicts.

Step 5: Install only the reviewed subset

After the source, selected names, target, collision report, licenses, copy mode, and rollback are reviewed, install only those names:

npx --yes skills@1.5.23 add /path/to/scientific-agent-skills \
  --skill scanpy anndata \
  --global --agent universal --yes --copy --full-depth

The upstream also documents gh skill install with a tag or SHA pin. Use it only if that extension is installed and its help output confirms the current syntax. Never turn an unavailable installer into an excuse to fall back to an unpinned branch or a partial SKILL.md-only copy.

Preserve the complete selected directory, except content that the license does not permit. Support files, scripts, templates, and local references are part of the reviewed unit. See installation and lifecycle.

Step 6: Keep dependencies isolated

The upstream collection intentionally spans incompatible scientific stacks. Do not install every package into one environment. For each selected skill:

1. read its pinned Python and system requirements;
2. create a project or task-specific environment;
3. preview package changes and trusted indexes;
4. verify GPU, compiler, Java, MATLAB, Conda, CUDA, or platform constraints;
5. record packages and versions actually installed;
6. keep credentials outside committed files and logs.

A successful skill-file install proves only discovery. It does not prove that its scientific runtime, dataset, API, model, or hardware path works.

Step 7: Reconfirm before scientific or external side effects

Require a separate reviewed scope before any of these actions:

• installing, upgrading, or removing packages, interpreters, drivers, or system tools;
• using API keys, tokens, service accounts, paid search, models, or databases;
• uploading private, patient, genomic, proprietary, or unpublished data;
• launching billed cloud, GPU, scheduler, or long-running autonomous jobs;
• controlling a robot, liquid handler, microscope, instrument, or laboratory platform;
• creating or changing ELN, LIMS, Benchling, DNAnexus, Latch, OMERO, protocols.io, or similar remote records;
• generating treatment, diagnosis, or patient-specific clinical guidance;
• submitting a manuscript, grant, protocol, report, or other external artifact.

Clinical and genomic skills are research aids. They do not diagnose or replace qualified professional judgment. Preserve uncertainty, provenance, and the human decision owner.

Step 8: Verify and report

For every installed selection, verify:

1. destination path and non-empty SKILL.md;
2. frontmatter name equals the directory;
3. support links resolve and no unexpected symlink escapes the folder;
4. installed bytes match the reviewed checkout;
5. exact source commit and selected license are recorded;
6. no unrelated local skill changed or disappeared;
7. one representative prompt selects the intended owner;
8. runtime checks are reported separately from instruction-file installation.

For refreshes, compare old and new pins before replacing anything. Re-run the license and risk audit because upstream triggers, scripts, dependencies, and terms can change independently.

Examples

Example 1: Audit the collection

Request: "K-Dense scientific-agent-skills 실제 목록과 라이선스부터 확인해줘."

Choose audit. Pin the checkout, run doctor, report the live tree and license exceptions, and stop before installation.

Example 2: Install a single-cell lane

Request: "K-Dense 팩에서 Scanpy와 AnnData만 프로젝트에 넣어줘."

Choose install. Plan scanpy and anndata, inspect their BSD licenses and runtime requirements, check collisions, and install only after the exact target is reviewed.

Example 3: Block restricted document copies

Request: "그 저장소의 docx, pdf, pptx, xlsx를 우리 카탈로그에 복사해줘."

Do not copy or derive them. Name the bundled Anthropic restrictions and route the requested file work to the existing local document skills.

Example 4: Gate laboratory execution

Request: "opentrons-integration을 깔고 이 프로토콜을 로봇에서 바로 돌려."

Separate installation from operation. Confirm robot model, deck layout, labware, liquids, simulation, credentials, physical supervision, abort path, and explicit execution approval before any live command.

Example 5: Route a generic paper request away

Request: "논문 주제 조사부터 원고와 리뷰 대응까지 끝내줘."

Use academic-research, not this pack wrapper, unless the user explicitly asks to source one named K-Dense skill.

Best practices

1. Treat the real pinned skills/ tree as the inventory authority.
2. Prefer one narrow owner and a pinned subset over a full bundle.
3. Inspect each selected license; repository-level MIT is not universal.
4. Never copy or adapt the restricted Anthropic document folders.
5. Rewrite broad activation language when adapting any permissible workflow so it does not compete with the whole local catalog.
6. Treat upstream security reports as review queues, not safety certificates.
7. Review remote-authority claims and untrusted web/data ingestion before use.
8. Keep scientific dependencies isolated by project and interpreter.
9. Keep secrets, patient data, private datasets, and unpublished results out of prompts, logs, commits, and public services unless explicitly authorized.
10. Separate installation evidence from scientific validity and external action evidence.
11. Re-audit moving upstream content before every refresh.
12. Preserve attribution, source commit, and rollback information.

References

• Pinned catalog and routing
• Installation and lifecycle
• Source audit and risk
• K-Dense-AI/scientific-agent-skills
• Pinned upstream commit
• Agent Skills specification