claude-code-expert: A Skill That Teaches Claude Code to Use Itself Better

32,000+ stars and zero gained in the last seven days. That's not a red flag — that's a skill that already saturated its target audience and is coasting on a stable install base. The "breakout" tag on SkillsMP is a bit misleading at this point, but the raw star count tells you this one has been around long enough to earn trust. I installed it, read the full SKILL.md, and spent some time with it in real sessions. Here's what I actually think.

What This Skill Does

At its core, claude-code-expert is a meta-skill. It gives Claude Code a deep, structured knowledge base about... Claude Code itself. The SKILL.md is essentially a comprehensive operating manual — covering CLAUDE.md hierarchy, hooks configuration, MCP server setup, sub-agent orchestration, memory file management, and CLI flags — baked into a skill that Claude can reference during any session.

The pitch is simple: instead of you explaining how hooks work every time you want to set one up, or pasting MCP documentation into the context window, this skill pre-loads that knowledge. You ask Claude to "add a PostToolUse hook that logs every Bash command," and it already knows the settings.json schema, the matcher syntax, and where the file lives.

It's written primarily in Portuguese (Brazilian), which is worth noting upfront. The technical content is solid and the code examples are language-agnostic, but if you're not comfortable with Portuguese prose, some of the explanatory sections will require a quick translate pass.

Why This Matters

Here's the real problem this skill addresses: Claude Code is a surprisingly deep tool, and its own documentation is scattered. The official Anthropic docs cover the basics, but things like the CLAUDE.md hierarchy across global/project/subfolder scopes, the exact JSON structure for hooks, or how to wire up a custom MCP server — these require you to either memorize the details or go hunting every time.

I've personally wasted 20-minute sessions just getting hooks syntax right. The PreToolUse vs PostToolUse distinction, the matcher field behavior, the difference between global and project-level settings.json — it's all documented somewhere, but not in one place, and certainly not loaded into Claude's context by default.

This skill consolidates that institutional knowledge into something Claude can actively use. That's a real gap being filled.

Key Capabilities Worth Calling Out

1. Hooks coverage is genuinely good. The skill includes working JSON examples for Stop hooks (the "beep when done" pattern is legitimately useful for long-running tasks), PostToolUse logging, and PreToolUse security scanning. The secret scanner hook example that runs a Python script before Bash commands is the kind of practical, production-ready pattern that takes experience to arrive at. This alone is worth the install for teams that want lightweight audit trails.

2. CLAUDE.md structure guidance. The recommended CLAUDE.md template is opinionated in a good way. The "Protocolo Pre-Tarefa" section — ensuring Claude always routes through an orchestrator before responding — is a pattern I've seen work well in complex projects. The suggestion to include an "Erros Conhecidos" (Known Errors) section with pre-baked solutions is something I'm now adding to my own projects.

3. MCP server reference. The MCP commands table and the custom Node.js MCP server skeleton are solid reference material. It's not exhaustive, but it covers the 80% case: filesystem, GitHub, postgres, sqlite, puppeteer. The custom server example is minimal but correct — enough to get started without wading through the full MCP SDK docs.

4. Memory file architecture. The ~/.claude/projects/<hash>/memory/ structure with MEMORY.md as an index file is a pattern I hadn't formalized before. The skill gives you a concrete directory layout and naming conventions. Whether you adopt it wholesale or adapt it, having a suggested structure beats inventing one from scratch mid-project.

5. CLI flags reference. The flags section is a quick-reference cheat sheet. --output-format json for scripting, --max-turns for autonomous workflows, --allowed-tools for sandboxed execution — these are the flags that actually matter in daily use, not the full --help dump.

Who Should Install This

Install it if: - You're new to Claude Code and want to skip the "read all the docs" phase - You regularly configure hooks or MCPs and want Claude to just know the syntax - You work on multiple projects and want consistent CLAUDE.md structure across all of them - You're onboarding teammates to Claude Code and want a shared knowledge baseline - You use Claude Code for automation scripts and need the CLI flags at your fingertips

Skip it if: - You've been using Claude Code for 6+ months and already have your own patterns locked in — you probably know everything in here already - You need English-only documentation in your skills (the Portuguese prose will slow you down) - You're looking for a skill that does things rather than one that knows things — this is reference material, not an action-oriented agent - Your project is simple enough that a single CLAUDE.md handles everything

How to Install

The skill lives in the Antigravity Awesome Skills repo. Install via npx:

npx antigravity-awesome-skills

Then select claude-code-expert from the skill browser, or install manually by dropping the SKILL.md into your skills directory:

# Global (available in all projects)
mkdir -p ~/.claude/skills
curl -o ~/.claude/skills/claude-code-expert.md \
  https://raw.githubusercontent.com/sickn33/antigravity-awesome-skills/main/skills/claude-code-expert/SKILL.md

# Per-project
mkdir -p .claude/skills
# same curl, different destination

Once installed, activate it in a session with /skill claude-code-expert or reference it in your CLAUDE.md.

Concerns and Limitations

The language barrier is real. About 60% of the prose is in Portuguese. The code examples are fine, but if you're trying to understand the reasoning behind a recommendation, you're either translating or guessing. For an English-speaking team, this creates friction. I'd love to see an English variant in the repo.

It's a knowledge dump, not a smart agent. This skill doesn't validate your hooks JSON, doesn't check if your MCP is actually running, and doesn't adapt to your specific project setup. It's a well-organized reference document that Claude can reason from — not an active assistant that does the work for you. Manage your expectations accordingly.

The curl-pipe-bash security allowlist comment is unexplained. There's a <!-- security-allowlist: curl-pipe-bash --> comment in the SKILL.md header. This is a non-trivial thing to silently allow in a skill, and there's no explanation of why it's there or what it enables. I'd want to understand that before deploying this in any environment where I care about security posture. It may be a registry artifact rather than an active instruction, but it deserves scrutiny.

Some content references external scripts that don't exist in this skill. The context_manager.py save command and context-guardian skill references assume you have other pieces of the Antigravity ecosystem installed. If you're installing this in isolation, those sections are dead ends.

The star count is inflated by ecosystem bundling. The 32k stars belong to the entire antigravity-awesome-skills repo, not this specific skill. It's a meaningful signal about the repo's health, but don't mistake it for 32k people specifically validating this skill.

Verdict

Install it. The hooks examples alone are worth it, and the CLAUDE.md structure guidance is solid enough to improve how most developers are currently using Claude Code. The Portuguese prose is a friction point but not a dealbreaker — the code examples are clear and the structure is logical enough to navigate.

Just be clear about what you're getting: this is a reference skill, not an action skill. It makes Claude smarter about Claude Code's own feature set. If you've already internalized all of this through months of use, you'll get less value. If you're still in the "where does that settings.json go again?" phase, this will save you real time.

I'd rate it a solid install for anyone in the first year of serious Claude Code usage, with the caveat that the curl-pipe-bash allowlist deserves a closer look before you deploy it in sensitive environments.

Links: - SkillsMP page - GitHub source