Claude code content validator

Claude Code Content Validator

Teaching content about a fast-moving tool rots quietly. A lesson that was correct when written can, months later, tell a learner to run a command that no longer exists, install a package that has been archived, or follow a file convention that has changed. The reader hits a wall and loses trust in the whole course.

This skill keeps Claude Code teaching content honest against the source of truth — Anthropic's official documentation at https://code.claude.com/docs — and proposes the fixes. It is the companion to writing the content: write it once, re-run this whenever the product moves or before you re-share.

This is different from a docs-vs-code sync. This skill validates content about Claude Code against Anthropic's external, live docs, not a repo against its own code.

When to use it

• Before publishing or re-sharing a Claude Code course, tutorial, or internal guide.
• After a Claude Code release, when commands, conventions, or models may have changed.
• Periodically (e.g. quarterly) on any "evergreen" Claude Code training material.

What to check (against the live docs — never from memory)

Read the content (markdown lessons, docs, slides-as-text), then verify every Claude Code claim by fetching the relevant page under https://code.claude.com/docs. Common drift points, with the facts to confirm each run:

• Slash commands are real and current — e.g. /clear, /compact, /context, /memory, /skills, /agents, /init, /help, /mcp, plan mode (Shift+Tab). Flag any invented command. (A frequent one: there is no /agent <name> — subagents auto-delegate by description or are asked for by name, and are managed with /agents.)
• Skills convention: a folder .claude/skills/<name>/SKILL.md.
• Subagents convention: a single file .claude/agents/<name>.md with name / description frontmatter — not a per-agent folder containing an AGENT.md.
• Hooks: configured in settings.json; many events exist (PreToolUse, PostToolUse, UserPromptSubmit, Stop, Notification, and more). Don't present a small fixed number as the complete set. Tool matchers use real tool names (Read, Write, Edit), not invented ones like WriteFileTool.
• MCP: added with claude mcp add (or by asking Claude); configured under mcpServers. Only name @modelcontextprotocol/server-* packages that are still active reference servers (filesystem, fetch, git, memory, sequentialthinking, time, everything); others (slack, github, gdrive, postgres, sqlite…) are archived or vendor-maintained — point to the registry / vendor instead of a dead package name. Re-verify each run.
• Settings: settings.json, permissions.allow / permissions.deny, exact tool names.
• Memory: project CLAUDE.md (./CLAUDE.md or .claude/CLAUDE.md) and user ~/.claude/CLAUDE.md; /init to generate.
• CLI flags used in headless examples actually exist (e.g. there is no --cwd; use --add-dir or cd into the directory).
• Model ids: no superseded ids — check the current latest before flagging.

Quality pass — garbage out, value in

• Flag filler, duplicated guidance, and dead/orphan files (e.g. content not reachable by the site's own router).
• Flag genuine gaps where one accurate tip would materially help a learner.
• Respect the content's audience. If the material is deliberately no-code (for business users), do not inject developer concepts (token caching, exit codes, env-var plumbing) just because the docs mention them. Match the existing altitude.

Respect the project's own rules

Before changing anything, read the project's CLAUDE.md / style guide and obey it — especially scope rules (e.g. "this course is closed: corrections yes, new lessons no") and voice. Corrections, factual fixes, and removing dead content are safe. Anything that changes scope is a proposal for the author, not an edit you make.

Output

1. A section-by-section report: ✅ valid / ⚠️ needs fix, each with the specific issue and the official source URL.
2. A proposed-edits list (file + change), grouped: factual fixes · removals (garbage/dead) · useful additions.
3. Make the safe corrections (factual fixes, dead-content removal, title/heading sync). For anything touching scope or judgment, list it for the author to approve.
4. Verify the content still builds/renders if it's a live site, then summarise what changed.

The honesty bar (non-negotiable)

Never validate a Claude Code fact from memory — the product moves faster than any model's training. Fetch the doc, cite the URL, and if a page can't be reached, say so rather than guessing. A report that says "couldn't confirm /foo, verify manually" is worth more than a confident wrong tick.