r-package-skill

R Package Skill Creation

Overview

Generates a new r-{package} skill from an R package's documentation. Core loop: capture intent → gather docs → draft skill → run test cases (with-skill + baseline in parallel) → grade → iterate until pass_rate >= 90% and no improvement for 2 iterations → optimize description.

When NOT to Use

• Package is simple/well-known (tidyverse core, base R)
• One-off usage -- just read the help
• A skill already exists that references this package

Ask Where to Install

Always ask before creating files. No silent defaults. If the user already specified a path, use it. See references/installation-paths.md for the prompt and agent-specific paths.

R-Specific Skill Structure

{install-path}/r-{package}/
  SKILL.md              # <500 words
  references/
    API.md              # REQUIRED: Complete CRAN reference manual
    vignette-name.md    # Include all CRAN vignettes

• references/API.md is required for every package skill.
• Include every CRAN vignette as references/{vignette}.md.
• Each generated SKILL.md must tell its reader: "Read references/API.md before writing code."
• Quick Reference table must show ALL important parameters, including optional/advanced. Mark tiers: required (no mark), (opt) optional, (adv) advanced. Never hide parameters that affect performance or output size (zoom levels, batch sizes, etc.).

Workflow

1. Capture intent. Ask what the user needs the skill to do, edge cases, input/output formats, and what makes it a success. Don't fetch docs until this is clear.
2. Gather docs. Fetch CRAN reference, vignettes, and (if available) btw tools. See references/doc-gathering.md.
3. Draft. Write SKILL.md, references/API.md, and vignettes.
4. Test. Spawn with-skill and baseline subagents in the same turn (parallel, not sequential). Grade outputs. Aggregate into benchmark.json. See references/testing.md.
5. Iterate. If pass_rate < 90% or it's still improving, analyze failures, edit the skill, rerun. Remove guidance transcripts show agents ignoring (YAGNI). Stop when pass_rate >= 90% AND no improvement for 2 iterations.
6. Optimize the description last. See references/description-optimization.md.

Description Recognition Tokens

Descriptions MUST include library({pkg}) and {pkg}:: tokens plus file-extension and domain triggers. Without the package-name tokens, descriptions that read as action-oriented ("Use when creating interactive maps") miss user prompts that contain library(mapgl) or mapgl::.

# Good
description: Use when code loads or uses freestiler, working with .pmtiles files, or preparing tiles for mapgl/MapLibre in R

# Bad: no recognition tokens
description: Use when creating PMTiles vector tilesets from large spatial datasets

See references/description-optimization.md for the train/held-out test method.

R Validators

Domain assertions call R validators in lib/r-validators/ at repo root via Rscript:

• plot-validator.R -- ggplot2/mapgl visualizations
• spatial-validator.R -- sf/spatial operations
• html-validator.R -- flextable/Shiny outputs
• numerical-validator.R -- collapse/regression results

Each returns JSON (valid, message, domain fields) for use as grading evidence.

R Execution Patterns

• Default: Rscript -e "code" or mcptools MCP.
• Create script files only if code is long-running or the user asks.
• If unavoidable: prefix temp_*.R or use tempfile(), clean up with on.exit(unlink(...)) or file.remove().

Common Mistakes

Mistake	Fix
Missing references/API.md	REQUIRED. Extract from CRAN reference manual.
Missing vignettes	Include every CRAN vignette as references/*.md.
Quick Reference hides parameters	Show ALL params that affect performance/output. Mark optional with (opt).
Running with-skill first, baseline later	Spawn both in the same turn so timing is comparable.
Assertion schema drift	Use text / passed / evidence in grading.json, not name / met / details.
Description too generic	Include library({pkg}) and {pkg}:: tokens.
Skipping baseline	Without a baseline, you can't tell whether the skill helped.
Optimizing description before skill works	Fix functionality first; tune triggering last.