Skill Writing Guide

Skill Writing Conventions

Structure checklist

Every skill needs:

• A folder with a descriptive kebab-case name (e.g., r-empirical-finance)
• A SKILL.md file inside that folder
• YAML frontmatter with name and description fields

Optional supporting files go in subfolders:

• scripts/ — code Claude can execute (R, Python, shell)
• references/ — documentation Claude reads into context as needed
• assets/ — templates, fonts, or other files used in output

Writing the description

The description is the single most important line. Claude uses it to decide

whether to load the skill. Claude tends to under-trigger, so be explicit:

Weak:

description: Helps with R code.

Strong:

description: >
  Conventions for writing empirical finance R code with data.table, fixest,
  arrow, and ggplot2. Use this skill whenever writing, reviewing, or debugging
  R scripts for panel data, regressions, or data pipelines — even if the user
  doesn't explicitly mention these packages.

Rules of thumb:

• Keep under 200 characters if possible, but clarity beats brevity
• List the phrases a user would actually type that should activate this skill
• Include "even if the user doesn't explicitly mention X" language
• Name the specific packages, tools, or frameworks involved

Writing the body

Use examples, not just rules

Claude learns patterns from examples more reliably than from abstract

instructions. Always show what you want, not just describe it.

Weak:

Use proper regression syntax.

Strong:

## Panel regressions
est <- feols(y ~ x1 + x2 | firm_id + year, data = dt, vcov = "cluster")

Encode your corrections

If you've corrected Claude 3+ times on the same thing across sessions,

that correction belongs in a skill. Common examples:

• "Never use tidyverse, always use data.table"
• "Always compile with pdfLaTeX, never XeLaTeX"
• "Always cluster standard errors on panel regressions"

Use clear constraint language

• "Always" and "never" for hard constraints
• "Prefer" and "avoid" for soft preferences
• "Warn if" for things Claude should flag but not block

Include "what NOT to do"

Negative instructions are surprisingly effective when Claude has known

defaults you want to override:

Do NOT use dplyr or any tidyverse package for data manipulation.
Do NOT use lm() for panel regressions — always use fixest::feols().
Do NOT use fontspec — it is incompatible with pdfLaTeX.

When to split a skill

• If a SKILL.md exceeds ~200 lines, move reference material into

a references/ subfolder and point to it from SKILL.md

• If two workflows only sometimes overlap, they should be two separate skills
• If a skill tries to cover both "how to write code" and "how to review code",

those are different tasks and deserve separate skills

Iteration workflow

• Write the skill (start small — one page is fine)
• Test with a natural prompt (not "use skill X to..." — test with

what you'd actually type, like "write me a DiD regression script")

• Note what Claude gets wrong or ignores
• Update the skill with corrections
• Repeat until the skill is stable across 3-4 different prompts
• Commit and push to the repo