Official Claude Reference Skill
Purpose
Ground any claim about Claude Code, skills, plugins, or the Claude API in Anthropic's official repositories and documentation rather than recall.
When to use
Use this skill whenever a question concerns Claude Code configuration, skill or plugin authoring, hooks, MCP wiring, model identifiers, pricing, or SDK behavior — including when the answer feels obvious.
Inputs
- the question or the feature being built
- the installed Claude Code version and platform
- which official sources are already available locally
Output
Return:
- the answer, with the specific file or document it came from
- the verification actually run — the command executed and its output
- version caveats where behavior differs between releases
- an explicit statement when the official source does not cover the question
Constraints
- check locally installed sources first —
~/.claude/plugins/marketplaces/holds cloned official marketplaces and their realSKILL.mdfiles, which beat any summary - treat
anthropics/claude-plugins-officialandanthropics/skillsas canonical for plugin and skill structure - verify CLI behavior by running
claude <command> --helprather than recalling flags - never state a model id, price, context limit, or rate limit from memory; look it up
- when local sources and memory disagree, the local source wins and the disagreement is worth reporting
- distinguish what is documented from what is observed behavior; observed behavior can change without notice
- a skill only auto-loads from
skills/<name>/SKILL.mdwith YAML frontmatter — verify layout against a real official skill before assuming a file will be discovered - if a claim cannot be grounded, say so rather than presenting a plausible answer as verified
Examples
- Confirm the exact frontmatter fields a skill supports before authoring one
- Check whether a Claude Code flag exists in the installed version
- Verify a model identifier and its context window before writing it into code