R/Pharma 2026 Workshop — AI Skills for Clinical Trial Design & ADaM Derivation

A hands-on workshop, to be delivered at R/Pharma 2026, on using Claude Code skills to accelerate two common clinical-programming tasks:

1. Deriving ADaM analysis datasets (ADSL, ADAE, ADVS/ADLB) from SDTM with the {admiral} R package.
2. Designing group sequential trials for survival endpoints — boundaries, sample size, multiplicity, and simulation-based verification.

Slides:

What you’ll learn

• What an agent skill is (a SKILL.md workflow plus supporting references and scripts) and how Claude Code picks one up.
• How the admiral skills encode function-selection logic and CDISC conventions so the agent writes QC-ready, submission-traceable R — with # REVIEW: markers at every protocol-specific decision.
• How to go from a trial scenario to a fully specified group sequential design and a Word report, with the design verified by lrstat simulation.
• How to run these skills yourself: install once, then drive them by describing the task in natural language.

You will clone this repo and run the skills live during the session.

Included skills

Setup

1. Clone

git clone https://github.com/jeffreyad/rpharma-2026-workshop.git
cd rpharma-2026-workshop

2. Install the dependencies

Use R >= 4.2 (the skills were exercised on R 4.4.1) and, for the group-sequential-design report step, Python >= 3.8.

Rscript install_packages.R                  # the 12 R packages
python3 -m pip install --user python-docx   # Word-report dependency

3. Verify

Rscript check_setup.R

Every line should read [ OK ]. If anything fails — a read-only R library, a CRAN mirror prompt, a compile error, a Python that pip and R disagree about — see INSTALL.md for the manual commands and per-platform fixes.

4. Point Claude Code at the skills

Option A — use this repo as your project. Open this folder in Claude Code. The root CLAUDE.md tells the agent where each skill lives and asks you to set your local Rscript path (needed by the GSD skill, which executes R). Then just describe a task:

Derive ADSL from the SDTM DM, EX, and DS domains for this study.

Design a Phase 3 trial for first-line metastatic NSCLC with co-primary PFS and OS, 1:1 randomization, target HR 0.70 for OS.

Option B — install into another project. Copy (or symlink) a skill folder into that project’s agent skills directory, e.g. for Claude Code:

cp -r skills/group-sequential-design ~/.claude/skills/
cp -r skills/admiral                 ~/.claude/skills/
# or symlink to keep them updated with `git pull`
ln -s "$(pwd)/skills/admiral" ~/.claude/skills/admiral

Option C — install from the upstream project with the skills CLI:

npx skills add RConsortium/pharma-skills --skill admiral-adsl --skill group-sequential-design

Credits

Every skill here is copied — without changes to its derivation logic — from the R Consortium pharma-skills project, part of the R Consortium Submissions Working Group Pilot 7 (a joint effort with the BBSW AI committee).

Each skill directory retains its original LICENSE file. See LICENSE for the repository license (MIT).

Go deeper after the workshop

• Source project: https://github.com/RConsortium/pharma-skills — full catalog (also sdtm-oak, clinical-trial-simulation, statistical-reviewer, rounding, and more), the skill development lifecycle, and community benchmarking.
• admiral: https://pharmaverse.github.io/admiral/ · CDISC ADaMIG v1.3
• gsDesign: https://cran.r-project.org/package=gsDesign · gsDesign2: https://cran.r-project.org/package=gsDesign2 · lrstat: https://cran.r-project.org/package=lrstat
• Submissions Working Group: https://rconsortium.github.io/submissions-wg/

Disclaimer

These skills produce drafting aids, not validated deliverables. LLM output can be wrong or vary between runs — always independently verify generated code and results against your own QC process before they inform any regulatory, clinical, safety, or GxP-regulated activity, and comply with your organization’s policies. The skills bundle scripts that execute in your environment; review a skill’s contents before running it and only supply data you are permitted to use. Provided as-is under the MIT License, with no warranty or endorsement from the R Consortium, the BBSW AI committee, or any contributor’s employer.