project-template

Starter repository for FerAl Lab decision-analysis models in R, following the DARTH coding framework (Alarid-Escudero et al., PharmacoEconomics 2019; reference implementation darthpack).

Coding and naming rules are not repeated here. They live in the lab-conventions skill of the decision-modeling plugin.

Start a new model

1. On GitHub, click Use this template → Create a new repository. Choose Private for anything that will touch restricted data or unpublished results.
2. Clone it and rename project-template.Rproj to <your repo name>.Rproj.
3. In DESCRIPTION, set Package, Title, Authors@R and Description.
4. Replace the first paragraph of CLAUDE.md with a description of the model.
5. Decide on a license for the new project and replace LICENSE. The MIT license here covers this template.
6. Rewrite this README for the model.

The new repository starts with its own history. Nothing from other lab repositories comes with it.

Folder layout

From Table 1 of the DARTH paper. Each folder has a .gitkeep so it survives "Use this template"; delete the .gitkeep once the folder has real files.

Folder	Contents
data-raw	Raw data plus the .R scripts that read, process and save it to data/
data	Cleaned and processed input data used by the components
R	Function definitions, one script per component, plus dataset documentation
analysis	Scripts controlling the flow, one per component, plus _master.R
output	Calibration output, PSA datasets and other generated data files (git-ignored)
figs	Figures
tables	Tables for a publication or report
report	R Markdown describing the full analysis, knitted for publication or HTA submission
vignettes	R Markdown documenting how to use the functions and data of each component
tests	testthat unit tests in tests/testthat/

The repository is structured as an R package, so the model loads with devtools::load_all(".").

analysis/_master.R

analysis/_master.R runs every component in order. Each component script must also run independently of the others: from a fresh R session at the project root, without relying on objects left in memory by an earlier script. A script loads what it needs from data/ or output/.

Model repositories only

This layout is for decision-analysis model repositories. Do not apply it to lab operations (meeting notes, administration, planning) or to other kinds of projects.

Never commit

Do not commit any of the following, to a private or a public repository:

• Restricted data. Anything covered by a data use agreement, IRB protocol or confidentiality terms, including individual-level and linked data and anything with PHI. Restricted data must live outside any repository folder, ideally in Stanford-approved secure storage for its risk classification, and code should read it from there through a path set in .Renviron. A restricted/ folder is excluded by .gitignore and blocked for Claude (see below), but that is only a safety net for a file that ends up in the wrong place, not somewhere to keep data.
• Large model outputs. Calibration runs, PSA datasets, simulated cohorts and other generated files. output/ is git-ignored; regenerate outputs from the code. For a small output that must be versioned, add it on purpose with git add -f.
• .Renviron. It holds API keys, tokens and machine-specific paths.
• renv/library. It is rebuilt from renv.lock. Commit renv.lock, renv/activate.R and renv/settings.json only.

.gitignore covers all four, but it is only a backstop. Check git status before every commit. If something on this list is committed by mistake, deleting it in a later commit does not remove it from the history. Stop, do not push (or push again), and tell the PI right away.

Claude Code settings

.claude/settings.json is shared by everyone who opens the repository in Claude Code:

• Restricted folders. A Read deny rule, Read(/**/restricted/**), stops Claude's file tools from reading anything under a restricted/ folder at any depth. It also covers cat, head and similar commands that name the file. It does not stop a command that reads files without naming them, such as grep -r pattern ., or a script that opens files itself. The rule is only a safety net: restricted data must live outside any repository folder (see "Never commit").
• Lab plugin. When you trust the folder, Claude Code offers the decision-modeling plugin from FerAlLab-git/feral-lab-os.

Put personal overrides in .claude/settings.local.json, which is git-ignored.

License

The template is released under the MIT License. See LICENSE.