Output Style, CLAUDE.md, or Skill — Picking the Right Mechanism

A team lead lists three things they want to standardize across the team's Claude Code usage:

1. Every response from Claude, on every task, should be written in a terse, no-preamble style with bullet points instead of prose, regardless of what the task is.
2. Claude should always know that the team's backend uses pnpm (never npm) and that the legacy/ directory should never be touched without an explicit request.
3. There should be a reusable, on-demand procedure for "audit this PR's changed files for missing test coverage," that any engineer can trigger by name when they want it, without it running on every single task.

For each of the three needs, name the correct mechanism (output style, CLAUDE.md, or a skill/subagent) and explain why the other two mechanisms would be a worse fit.