# claude-skills-reference

Snapshot of Claude Code skills, copied here as a reference for Jerry to study how Anthropic structures task-specific guidance for a coding agent.

These are **not** Jerry's own skills — they're examples of the pattern. Each skill is a directory with a `SKILL.md` (instructions to the agent) plus optional supporting files (references, hooks, templates).

## What's here

| Skill | Source | What it does |
|---|---|---|
| `caveman/` | `caveman` plugin | Ultra-compressed communication mode — drops articles, filler, pleasantries. Three intensity levels (lite / full / ultra) plus three classical-Chinese variants. |
| `caveman-commit/` | `caveman` plugin | Generates conventional-commit messages with the same compression discipline. ≤50 char subject, body only when "why" isn't obvious. |
| `caveman-review/` | `caveman` plugin | Compressed PR review comments. One line each: location, problem, fix. |
| `compress/` | `caveman` plugin | Compresses natural-language memory files (CLAUDE.md, todos) into the caveman format. Saves input tokens. |
| `mermaid/` | Anthropic built-in | Generates Mermaid diagrams from prose requirements. Supports 18+ diagram types with per-type reference docs under `mermaid/references/`. |

## Pattern worth borrowing

Each skill has a top-level `SKILL.md` that:

1. States the trigger ("when user says X, or invokes /Y")
2. Defines the rules (what to do, what NOT to do)
3. Gives worked examples (before/after, good/bad)
4. Lists boundaries (when to deactivate / when not to apply)

The mermaid skill goes further: it has a per-diagram-type `references/` directory the agent loads on demand, so the SKILL.md stays small while the catalogue can be large.

For Jerry's review work, the same shape would help — one `code-review-skill/SKILL.md` listing the playbook rules, with `references/<topic>.md` files the agent loads when the review touches that topic (e.g. `references/go-concurrency.md`, `references/auth.md`).

## How Jerry could use these

- **Read `SKILL.md` files** to see how compact, behavior-changing instructions look. The caveman skills in particular cram a behavior change into ~50 lines without sacrificing precedence rules.
- **Steal the trigger pattern**: a one-line "when the user / agent does X, do Y" at the top of every skill makes the agent's behavior predictable.
- **Steal the "examples" pattern**: every skill has before/after examples. Concrete examples > abstract rules.

— Claude