What this resource does
Core uses
This repository is a large collection of separate scientific Skills
The reviewed source contains 163 scientific and research Skill files. This page groups a representative selection by task rather than listing the whole collection.
The component cards are a selection from the 163 Skill files in the reviewed snapshot.
Scientific literature and database access
Provides separate Skills for literature work and querying scientific databases in biology, chemistry, medicine, materials science, and related fields.
- Typical inputs
- Research question, entity, identifier, or search terms; Selected database or domain Skill
- Output
- Source records, literature results, or structured scientific metadata from the selected component.
Best for: Finding research evidence or retrieving domain-specific scientific information.
Skills for this task
database-lookup
Queries documented public database APIs with explicit endpoints, filters, pagination, and provenance.
literature-review
Conducts structured literature reviews across academic databases and creates Markdown or PDF review documents.
biopython
Supports sequence processing, biological file parsing, phylogenetics, and programmatic NCBI or PubMed access.
Suggested workflow requests (1)
Use available skills you have access to whenever possible. Work only with authorized synthetic or de-identified data. Parse the VCF with pysam, annotate variants with Ensembl VEP, retrieve ClinVar/COSMIC/NCBI Gene/UniProt evidence, and verify literature sources. Build an evidence- traceable research summary with scientific-writing. If clinical-reports is used, create only a visibly marked draft structure from a verified source-fact manifest for qualified review; do not diagnose, assess individual risk, recommend treatment, or determine trial eligibility.
From README.md.
Computational research analysis
Includes analysis workflows for genomics, chemistry, statistics, machine learning, geospatial science, and scientific simulation.
- Typical inputs
- Research dataset or domain inputs; Selected method and component requirements
- Output
- Component-dependent analysis results, code, figures, or reports.
Best for: Applying a documented scientific package or method to research data.
Skills for this task
scanpy
Runs established single-cell RNA-seq workflows including quality control, clustering, differential expression, and visualisation.
rdkit
Provides fine-grained cheminformatics workflows for molecular parsing, descriptors, fingerprints, similarity, and reactions.
pymc
Supports Bayesian modelling, MCMC, variational inference, model comparison, and posterior checks with PyMC.
geopandas
Works with geospatial vector data for spatial joins, overlays, coordinate transformations, and mapping.
Suggested workflow requests (5)
Use available skills you have access to whenever possible. Query ChEMBL for EGFR inhibitors (IC50 < 50nM), analyze structure-activity relationships with RDKit, generate improved analogs with datamol, perform virtual screening with DiffDock against AlphaFold EGFR structure, search PubMed for resistance mechanisms, check COSMIC for mutations, and create visualizations and a comprehensive report.
Use available skills you have access to whenever possible. Load 10X dataset with Scanpy, perform QC and doublet removal, integrate with Cellxgene Census data, identify cell types using NCBI Gene markers, run differential expression with PyDESeq2, infer gene regulatory networks with Arboreto, enrich pathways via Reactome/KEGG, and identify therapeutic targets with Open Targets.
Use available skills you have access to whenever possible. Analyze RNA-seq with PyDESeq2, process mass spec with pyOpenMS, integrate metabolites from HMDB/Metabolomics Workbench, map proteins to pathways (UniProt/KEGG), find interactions via STRING, correlate omics layers with statsmodels, build predictive model with scikit-learn, and search ClinicalTrials.gov for relevant trials.
Use available skills you have access to whenever possible. Retrieve AlphaFold structures, identify interaction interface with BioPython, search ZINC for allosteric candidates (MW 300-500, logP 2-4), filter with RDKit, dock with DiffDock, rank with DeepChem, check PubChem suppliers, search USPTO patents, and optimize leads with MedChem/molfeat.
Use available skills you have access to whenever possible. Query NCBI Gene for annotations, retrieve sequences from UniProt, identify interactions via STRING, map to Reactome/KEGG pathways, analyze topology with Torch Geometric, reconstruct GRNs with Arboreto, assess druggability with Open Targets, model with PyMC, visualize networks, and search GEO for similar patterns.
From README.md.
Research communication and laboratory work
Contains Skills for scientific writing, figures, presentations, protocols, electronic laboratory notebooks, and research-platform integrations.
- Typical inputs
- Draft text, figures, protocols, or platform task; Target deliverable and relevant constraints
- Output
- Draft research materials or platform-specific records produced by the selected Skill.
Best for: Preparing research outputs or working with a documented laboratory platform.
Skills for this task
scientific-writing
Drafts scientific manuscripts through an outline-first process with IMRAD structure and reporting-guideline support.
scientific-visualization
Coordinates publication-ready figures with multi-panel layouts, statistical annotations, and journal formatting.
benchling-integration
Connects to Benchling SDK and APIs for registry entities, inventory, ELN entries, workflows, and data queries.
protocolsio-integration
Reads, validates and exports protocols.io data, and prepares non-executing plans for write operations. Its bundled tools do not execute mutations.
Use boundaries
Limits and checks
Component variability
A collection-level description cannot establish the behaviour of every component.
Inspect and test the selected Skill before research use.
External services
Costs, limits, data transfer, or account terms may apply.
Check the selected component's environment variables and service terms.
Scientific validity
Unchecked output could lead to unsupported research conclusions.
Validate methods, code, calculations, and source records independently.
More boundaries
- Task boundary: The collection does not provide one shared workflow that combines every scientific Skill into a complete research project.
- Input boundary: A component cannot use restricted databases, local software, or credentials that the user has not legitimately provided.
- Decision boundary: Scientific interpretations, clinical implications, calculations, and citations require expert review against original evidence.
- Individual Skills may use licences different from the repository's MIT licence. Check the selected Skill's terms before use.
Technical details
- Resource type
- Skill Collection
- Author or maintainer
- K-Dense
- Version
- 2.64.0
- Source last updated
- 8 Jul 2026
- Last verified
- 13 Jul 2026
- Licence
- MIT
- Access
- Publicly available
- Additional costs
- Platform terms or usage limits may apply. API usage fees may apply for selected components. External services, software, compute, or data access may have separate costs.
- Skill instruction language
- English
- Documentation language
- English
- Repository languages
- Python, Markdown
- Dependencies
- Node.js and npx for the recommended installer; Component-specific Python packages, APIs, databases, and credentials
- Review status
- Quick test completed
Continue exploring
Related resources
Official package contents
Files
Read the official English instructions without leaving this profile. This view includes the English files represented by this profile. Non-English source files, supporting code, assets, tests, and references may not be shown.
Official package files
11 files available here
skills11
benchling-integration1
biopython1
database-lookup1
geopandas1
literature-review1
protocolsio-integration1
pymc1
rdkit1
scanpy1
scientific-visualization1
scientific-writing1
Read on this page
Select a file
Choose a file from the tree to read its official instructions without leaving this page.
Independent testing
Evaluation report
See exactly what was tested, which checks were used, and what the result does not cover.
Evaluations are mainly based on automated AI-assisted checks. Results are for reference only.
Evaluation coverage
3 components tested separately
Each score applies only to the named component, cases, application and model. The scores are not combined into a Repository rating.
8 named components were not tested
literature-reviewRequires a configured external API or service account. Not tested in this review. Contact us if you would like this capability tested.biopythonscanpyNot tested in this review. Contact us if you would like this capability tested.rdkitpymcgeopandasbenchling-integrationprotocolsio-integration
Tested component 1scientific-writing100 / 100 · One example
One example
Turning three research findings into a Results paragraph
The component returned a short Results paragraph. It kept the supplied numbers and study limitation, and added no group averages, p-values or extra confidence intervals.
Codex CLI · gpt-5.6-luna · 31 Aug 2026
What we asked it to do
Write a short Results section from three supplied research findings. Keep the numbers and limitations unchanged, and do not invent missing statistics.
Exact request used in the test
Draft a concise Results section from the supplied evidence outline, preserving all claim IDs, sample sizes, estimates, units, and limitations. Do not calculate, estimate, or invent group means, standard deviations, confidence intervals, p-values, or citations that are absent from the outline.
What was provided
Three synthetic findings, not results from a real study. C001-C003 label the claims; E001-E003 label the supplied evidence. The request explicitly asked the component to keep these labels.
Original input
Synthetic evidence outline: C001: 62 of 80 participants completed follow-up (E001). C002: the intervention group improved by 2.4 points relative to control, 95% CI 0.8 to 4.0 (E002). C003: the study was conducted at one institution and was not designed to establish causality (E003). No group means or additional confidence intervals are supplied.
Output from the Skill
The complete original paragraph is below. The bracketed labels link each sentence to the supplied findings; they are not published references. No wording has been rewritten for this report.
Results
Follow-up was completed by 62 of 80 participants [claim:C001] [evidence:E001]. The intervention group improved by 2.4 points relative to control, with a 95% CI of 0.8 to 4.0 [claim:C002] [evidence:E002]. The study was conducted at one institution and was not designed to establish causality [claim:C003] [evidence:E003]. No group means or additional confidence intervals were supplied.
Our assessment
The numbers stayed the same
The output keeps 62 of 80 participants, a 2.4-point difference and the supplied 95% interval of 0.8 to 4.0. It does not add averages, standard deviations or p-values.
The study limit stayed visible
The paragraph still says the study took place at one institution and was not designed to establish causality. It does not turn the reported difference into proof of cause and effect.
This is drafting, not fact checking
The component arranged the provided statements into prose. It did not verify a real dataset or inspect papers. The claim labels came from the test input.
How this example was scored
This score covers one selected example from 9 earlier cases, using the five existing weighted dimensions. It does not rate the whole component or Repository. The original output is unchanged. This is a review of retained output, not a new Skill run.
Five scoring dimensions
Passed checks receive 1 point, partly met checks 0.5, and failed checks 0. Checks outside this example are excluded, not counted as passes. The five dimension percentages are combined using the weights shown below.
Did it complete the task?
100.0%The output is a Results paragraph with the requested heading and three evidence-linked claims.
Did it run without problems?
100.0%The selected prose task returned a complete answer in the recorded setup.
Did it preserve and handle the supplied information responsibly?
100.0%The supplied numbers, uncertainty, identifiers and limitation were preserved. No new statistics were added.
Did it do the documented task in the tested setup?
100.0%Only this short drafting task in the recorded Codex setup is covered.
Were the setup and use instructions clear enough to follow?
100.0%The guide identifies the evidence inputs and the drafting workflow used here.
Limits of this example
- Only a short supplied outline was drafted. Longer manuscripts, reference verification, missing-evidence handling and reviewer reports are not covered by this score.
- The three-example August 28 writing report is historical. Its score is not combined with this selected August 31 case.
Tested component 2database-lookup100 / 100 · One example
One example
Looking up a human gene in NCBI Gene
The component returned all seven requested fields for TP53. Each value matches the database response saved during the test. This lookup used public access without setting up an API key.
Codex CLI · gpt-5.6-luna · 4 Sept 2026
What we asked it to do
Look up human gene 7157 in NCBI Gene and return its name, description, chromosome, location and organism identifiers.
Exact request used in the test
Query the documented NCBI Gene eSummary endpoint for Gene ID 7157 using unauthenticated public access. Return the UID, gene symbol, description, chromosome, map location, organism scientific name, and taxon ID. Include the exact endpoint, parameters, UTC access time, and a field-level comparison against the authoritative response.
What was provided
One database identifier (7157), the expected organism (human), and a list of seven fields to retrieve. The name TP53 and human identifiers were supplied as identity checks; the remaining fields had to come from the database.
Original input
{
"database": "NCBI Gene",
"gene_id": "7157",
"required_fields": ["uid", "name", "description", "chromosome", "maplocation", "organism.scientificname", "organism.taxid"],
"predefined_identity": {"uid": "7157", "name": "TP53", "organism.scientificname": "Homo sapiens", "organism.taxid": "9606"},
"access_mode": "unauthenticated public access",
"scope": "targeted lookup"
}Output from the Skill
The Results section below is copied from the actual answer. UID is the database record number, map location is the position on the chromosome, and taxon ID identifies the species. The full technical request log is not displayed.
uid: 7157
gene symbol: TP53
description: tumor protein p53
chromosome: 17
map location: 17p13.1
organism scientific name: Homo sapiens
taxon ID: 9606
Our assessment
Seven fields can be checked directly
The saved response contains record 7157, TP53, tumor protein p53, chromosome 17, location 17p13.1, Homo sapiens and species identifier 9606. All seven appear unchanged in the answer.
No account setup was needed for this lookup
The official NCBI reference allows access without an API key. The recorded request used that route. This does not establish access to the other databases supported by the component.
This was a single-record lookup
The test asked for one known human gene, not a gene discovery search or an exhaustive dataset. Other earlier attempts, including the aspirin lookup with an incomplete saved response, are not included in this example score.
How this example was scored
This score covers one selected example from 7 earlier cases, using the five existing weighted dimensions. It does not rate the whole component or Repository. The original output is unchanged. This is a review of retained output, not a new Skill run.
Five scoring dimensions
Passed checks receive 1 point, partly met checks 0.5, and failed checks 0. Checks outside this example are excluded, not counted as passes. The five dimension percentages are combined using the weights shown below.
Did it complete the task?
100.0%All seven requested fields were returned for the specified human gene.
Did it run without problems?
100.0%The recorded unauthenticated request returned HTTP 200 and a readable answer.
Did it preserve and handle the supplied information responsibly?
100.0%Each displayed field matches the saved database response. Missing-field behavior is outside this case.
Did it do the documented task in the tested setup?
100.0%Only one NCBI Gene lookup is covered, not the other databases in the collection.
Were the setup and use instructions clear enough to follow?
100.0%The official reference provides the endpoint, parameters and access conditions.
Limits of this example
- The values were checked against the response retained on 4 September 2026, not a new live query.
- Missing records, pagination, rate limits and other databases are outside this score.
Tested component 3scientific-visualization100 / 100 · One example
One example
Creating a two-panel figure from 24 supplied records
In this rerun, scientific-visualization produced both panels and saved PNG and PDF files. All 24 plotted means and all 48 interval limits match the supplied table. Long labels are visible, and the two treatments share one legend.
Codex interactive session · Model not recorded · 5 Sept 2026
What we asked it to do
Create a two-panel figure from the same 24-row table, comparing two treatments across three conditions. Show each supplied mean and interval, keep long labels readable, and export PNG and PDF at 180 by 120 mm.
Exact request used in the test
# Scientific Visualization rerun User request: rerun the scientific-visualization component on the same data. Task retained from the previous case: create a two-panel faceted figure comparing two treatments across three conditions, showing the supplied means and 95% intervals, replicate entries, long condition labels, shared units, one figure-level legend, redundant visual encodings and fixed 180 mm x 120 mm PNG/PDF exports. Retain all 24 rows unchanged. Execution: follow the retained official SKILL.md directly in the current Codex conversation, using its style and export helpers. This is a new interactive run, not execution or repair of the old renderer. The previous failure is known, so this is not a blinded or same-model replication. No external API configuration is needed. Input clarification: the CSV supplies two replicate-level means and intervals per combination. It does not supply individual-level raw observations. Plot the supplied entries individually and explain this limitation rather than fabricate raw samples or recompute group intervals. No target journal was supplied. Treat this as a general scientific figure, not a certified submission-ready figure.
Original data
A synthetic table with two measurement panels, three conditions and two treatments. Each combination has two replicate entries with a mean, interval limits and units. These 24 rows are unchanged from the earlier test; they are not individual participant data.
View the 24 supplied rows
| condition | panel | treatment | replicate | mean | lower_ci | upper_ci | unit |
|---|---|---|---|---|---|---|---|
| Short-term response after baseline correction | Panel A: efficacy | Standard care | 1 | 10.2 | 9.4 | 11.0 | response units |
| Short-term response after baseline correction | Panel A: efficacy | Standard care | 2 | 10.4 | 9.6 | 11.2 | response units |
| Short-term response after baseline correction | Panel A: efficacy | Intervention | 1 | 11.7 | 10.9 | 12.5 | response units |
| Short-term response after baseline correction | Panel A: efficacy | Intervention | 2 | 11.9 | 11.1 | 12.7 | response units |
| Short-term response after baseline correction | Panel B: tolerability | Standard care | 1 | 12.2 | 11.4 | 13.0 | response units |
| Short-term response after baseline correction | Panel B: tolerability | Standard care | 2 | 12.4 | 11.6 | 13.2 | response units |
| Short-term response after baseline correction | Panel B: tolerability | Intervention | 1 | 13.7 | 12.9 | 14.5 | response units |
| Short-term response after baseline correction | Panel B: tolerability | Intervention | 2 | 13.9 | 13.1 | 14.7 | response units |
| Long-term response after baseline correction | Panel A: efficacy | Standard care | 1 | 13.2 | 12.4 | 14.0 | response units |
| Long-term response after baseline correction | Panel A: efficacy | Standard care | 2 | 13.4 | 12.6 | 14.2 | response units |
| Long-term response after baseline correction | Panel A: efficacy | Intervention | 1 | 14.7 | 13.9 | 15.5 | response units |
| Long-term response after baseline correction | Panel A: efficacy | Intervention | 2 | 14.9 | 14.1 | 15.7 | response units |
| Long-term response after baseline correction | Panel B: tolerability | Standard care | 1 | 15.2 | 14.4 | 16.0 | response units |
| Long-term response after baseline correction | Panel B: tolerability | Standard care | 2 | 15.4 | 14.6 | 16.2 | response units |
| Long-term response after baseline correction | Panel B: tolerability | Intervention | 1 | 16.7 | 15.9 | 17.5 | response units |
| Long-term response after baseline correction | Panel B: tolerability | Intervention | 2 | 16.9 | 16.1 | 17.7 | response units |
| Adverse-event burden during follow-up | Panel A: efficacy | Standard care | 1 | 16.2 | 15.4 | 17.0 | response units |
| Adverse-event burden during follow-up | Panel A: efficacy | Standard care | 2 | 16.4 | 15.6 | 17.2 | response units |
| Adverse-event burden during follow-up | Panel A: efficacy | Intervention | 1 | 17.7 | 16.9 | 18.5 | response units |
| Adverse-event burden during follow-up | Panel A: efficacy | Intervention | 2 | 17.9 | 17.1 | 18.7 | response units |
| Adverse-event burden during follow-up | Panel B: tolerability | Standard care | 1 | 18.2 | 17.4 | 19.0 | response units |
| Adverse-event burden during follow-up | Panel B: tolerability | Standard care | 2 | 18.4 | 17.6 | 19.2 | response units |
| Adverse-event burden during follow-up | Panel B: tolerability | Intervention | 1 | 19.7 | 18.9 | 20.5 | response units |
| Adverse-event burden during follow-up | Panel B: tolerability | Intervention | 2 | 19.9 | 19.1 | 20.7 | response units |
Generated figure
This is the unchanged figure from the new interactive run. Circles show standard care; squares show the intervention. Filled and open points show the two supplied replicate entries. Horizontal lines show the supplied 95% intervals.
What the result shows
The two panels now show the intended measurements
The left panel shows efficacy and the right panel shows tolerability. Both contain the same three conditions. Blue circles and orange squares distinguish the treatments; filled and open markers distinguish the two supplied replicate entries.
Every supplied value is represented
We compared the plotted points and interval ends with the original table: all 24 means and all 48 interval limits matched. No row was dropped, averaged with another row or changed.
Both exports are readable
The PNG and one-page PDF show the same two panels. Long condition names wrap onto two lines without being cut off. The legend appears once, and both axes show response units. The lower notes are close to the axis labels but remain separate and readable.
The chart does not invent participant data
Each point represents a mean already supplied in the table. The caption explains that individual-level measurements were not provided. The displayed intervals were copied from the input, not calculated again.
How this example was scored
This score covers this rerun only, not the whole component or Repository. We followed the official instructions in the current Codex session after seeing the earlier failure. All applicable checks passed; missing-value and extreme-value handling were not tested. The score measures these checks, not aesthetic quality, first-attempt reliability or journal acceptance.
Five scoring dimensions
Passed checks receive 1 point, partly met checks 0.5, and failed checks 0. Checks outside this example are excluded, not counted as passes. The five dimension percentages are combined using the weights shown below.
Did it complete the task?
100.0%Both panels and both exports were produced. All plotted values match the input, with readable labels and one shared legend.
Did it run without problems?
100.0%This one rerun completed with the prepared local dependencies. This is not evidence of success on every attempt.
Did it preserve and handle the supplied information responsibly?
100.0%Values and intervals were preserved, with no invented raw observations. Missing and extreme values are outside this case.
Did it do the documented task in the tested setup?
100.0%The documented static-figure task worked in this recorded local setup. Other models and platforms were not tested.
Were the setup and use instructions clear enough to follow?
100.0%The official instructions identify the inputs, local tools, output formats and review steps used for this figure.
Limits of this example
- This was an assisted interactive rerun with the previous failure known, not an independent repeat under identical conditions.
- The exact AI model identifier was not recorded for this session. It must not be attributed to the model used in the earlier test.
- The input contains record means and interval limits, not individual participant measurements. No extra raw observations or new confidence intervals were generated.
- Missing values, extreme values, interactive charts, other export formats and journal-specific requirements were not tested.
Want to see more tests?
Tell us which module or task you would like us to test next.