Using gsDesign AI skills

Overview

The gsDesign repository includes lightweight AI workflow instructions for coding agents. These files are not statistical methods documentation. They are promptable development guides that help coding agents use consistent package maintenance habits: checking git status, updating NEWS.md, bumping development versions, running focused and broad tests, rebuilding pkgdown, and pushing issue branches.

The repository includes native instruction files for common coding agents:

• AGENTS.md for Codex-style agents.
• CLAUDE.md for Claude Code.
• GEMINI.md for Gemini CLI.
• .github/copilot-instructions.md for GitHub Copilot.
• .gemini/styleguide.md for Gemini Code Assist.

The installed R package also ships copies under inst/ai/, available with:

system.file("ai", package = "gsDesign")

This makes the instructions available to users who install the package without cloning the GitHub repository.

Installing instructions in a project

After installing gsDesign, you can copy an instruction file into a local project. For example, to add Codex-style instructions:

file.copy(
  from = system.file("ai", "AGENTS.md", package = "gsDesign"),
  to = "AGENTS.md",
  overwrite = FALSE
)

For GitHub Copilot:

dir.create(".github", showWarnings = FALSE)
file.copy(
  from = system.file("ai", "copilot-instructions.md", package = "gsDesign"),
  to = file.path(".github", "copilot-instructions.md"),
  overwrite = FALSE
)

For Gemini Code Assist:

dir.create(".gemini", showWarnings = FALSE)
file.copy(
  from = system.file("ai", "gemini-styleguide.md", package = "gsDesign"),
  to = file.path(".gemini", "styleguide.md"),
  overwrite = FALSE
)

Example prompt and response

An AI skill works best when the user gives a specific development task and lets the instruction file supply the package workflow details.

Example prompt:

Use the gsDesign workflow instructions. Fix issue #264 on the current branch: update simBinomialSeasonalExact() so simulated trials stop at efficacy or futility boundary crossings, update toInteger() for rare-event equal allocation, add regression tests, update NEWS.md, rebuild pkgdown, run the package checks, commit, and push if everything is OK.

Example response:

I checked the branch and worktree, implemented the simulation stopping and event-rounding updates, added focused regression tests, bumped the package to the next development version, updated NEWS.md, rebuilt pkgdown, and ran R CMD check --no-manual. The only remaining check warning is a local macOS/R toolchain warning from R_ext/Boolean.h, not package code. I committed the changes and pushed the issue branch to origin.

The response is useful because it reports the workflow steps the agent took and separates package-code results from local toolchain noise.

What these instructions do not do

AI workflow instructions do not change the statistical behavior of gsDesign. They also do not replace human review. They help make agent-assisted package maintenance more reproducible by placing the expected repository workflow where coding agents can find it.