UK Biobank Data Processing and Survival Analysis Toolkit.

UKBAnalytica Skill Pack (UKBAnalytica_skills)

Agent-runtime-agnostic skill bundle for the UKBAnalytica R package. The pack works with any agent stack that supports a "skill" / "tool description" loaded from YAML front-matter + Markdown (Claude Code skills, OpenAI Assistants instructions / function-tool descriptions, custom RAG-driven agents, etc.).

The pack lives at inst/skills/UKBAnalytica_skills/ in the repository root. Load it via file.path(getwd(), "inst/skills/UKBAnalytica_skills") or see INSTALL.md for per-runtime loader snippets.

Charter (read first)

All skills in this pack enforce the following non-negotiable rules:

1. Script-generation boundary. These skills help a local agent generate scripts, analysis plans, package-usage guidance, and manuscript text from aggregate outputs. They are not permission for an agent to read or process real UK Biobank RAP participant-level data.
2. RAP-only execution. Scripts that touch real UK Biobank data must be run by the user inside the approved UK Biobank Research Analysis Platform (RAP), typically RAP JupyterLab → Terminal → R.
3. No real rows in agent context. Do not give the agent participant-level files, R objects, screenshots, logs, tracebacks, row samples, head(data), eid, exact dates, raw RAP fields, per-row predictions, or row-level SHAP matrices. This applies even when identifiers have been removed.
4. Aggregate outputs only. The user may share aggregate summaries with the agent, including participant-flow counts, baseline tables, regression summaries, model metrics, enrichment results, and rendered figures.
5. Schema-only prompts. The user may describe column names, variable roles, and intended analyses. The agent uses synthetic toy data for smoke tests and writes final scripts for RAP execution.
6. Large extracts go async. When the requested field count is large, skills route the user through rap_submit_extract() (DNAnexus table-exporter) rather than pulling everything into the R session.
7. UKBAnalytica is the sole executor. Skills wrap real exported functions only — every function name in a skill must be present in NAMESPACE.
8. No journal-brand styling. Plotting guidance uses neutral palette names (ukbsci_clinical, ukbsci_diverging, ukbsci_sequential) and avoids referencing specific top-tier journal brands.

Provider compatibility

The front-matter is intentionally minimal (name, description) so it parses identically across providers. No provider-specific keys are used. Regardless of provider, these skills must not be used to send real participant-level RAP data to the agent.

Pack layout

inst/skills/UKBAnalytica_skills/
├── README.md                       ← this file
├── INSTALL.md                      ← per-provider install snippets
├── MANIFEST.json                   ← machine-readable index (name, description, path)
├── ukbsci-rap-extract/             (P2) RAP discover / plan / extract
├── ukbsci-cohort/                  (P2) disease definitions + survival cohort
├── ukbsci-workflow/                (P2) end-to-end orchestrator
├── ukbsci-regression/              (P3) batch lm / logit / Cox + extensions
├── ukbsci-survival/                (P3) KM + risk table + log-rank
├── ukbsci-baseline/                (P3) tableone Table 1
├── ukbsci-propensity/              (P4) PS / PSM / IPTW / balance
├── ukbsci-mediation/               (P4) regmedint 4-way decomposition
├── ukbsci-subgroup-sensitivity/    (P4) subgroup × interaction + sensitivity
├── ukbsci-imputation/              (P4) mice + Rubin pooling
├── ukbsci-proteomics/              (P5) Olink / STRING / GO / KEGG / PPI
├── ukbsci-ml/                      (P5) classification + survival ML + SHAP
├── ukbsci-preprocess/              (P5) variable cleaning + composites
└── ukbsci-plot/                    (P6) forest / volcano / calibration / theme

Each skill directory contains:

SKILL.md           ← YAML front-matter (name, description) + body
README.md          ← human-readable overview
references/
  functions.md     ← every exported function signature + caveats
  rap-guardrails.md← what is forbidden in this module
  examples.md      ← copy-pastable minimal examples
evals/
  evals.json       ← trigger-recall test cases for the skill router

Trigger-word routing table

Every SKILL.md description ends with the phrase "UK Biobank RAP" or "UKBAnalytica" to keep the router from confusing these skills with generic R / statistics skills. Triggers below are written into the description field verbatim.

Status

All 14 skills shipped (v1.0.0). Phase grouping reflects the writing order:

See supp/UKBAnalytica-skill-roadmap.md for the full implementation roadmap.

Citation

When agents emit R scripts based on this pack, the recommended citation header at the top of every generated script is:

###############################################################################
# UKBAnalytica Citation:
# He N. UKBAnalytica: Scalable Phenotyping and Statistical Pipeline for
# UK Biobank RAP Data. R package version 1.0.0.
# https://github.com/Hinna0818/UKBAnalytica
###############################################################################