bot_vault

Claude Code Daily Practice: Plugins, Skills, and Config

2 min read

Summary

A practitioner’s guide distilled from daily Claude Code use, covering CLAUDE.md authoring, settings security, the plugins ecosystem, skill installation, MCP server choices, and terminal customization. Most advice applies to other coding agents (Codex, Gemini CLI) as well.

每日使用 Claude Code 的實踐指南，涵蓋 CLAUDE.md 設定、安全設定、插件生態系統、技能安裝和 MCP 伺服器選擇。大多數建議也適用於其他 coding agents（Codex、Gemini CLI）。

Key Points

  • Global ~/.claude/CLAUDE.md should only hold personal preferences; don’t repeat what’s in the system prompt — ask it first
  • Per-file-type rules belong in ~/.claude/rules/ so they only load for relevant file types
  • Use a "deny" list in settings.json to block dangerous commands, but also add a Bash hook (guard-bash-paths.py) since Read deny alone doesn’t stop bash access
  • Prefer CLI tools over Playwright/GitHub MCPs — use playwright-cli/agent-browser skills and gh command instead
  • Codex MCP is ironically useful: other agents (including Claude) can use Codex via MCP for multi-model opinions
  • /magi-ex and /second-opinions skills give parallel reviews from multiple model families (Opus, Codex, Gemini)

Insights

The “Gotchas” section pattern is the highest-signal content in any CLAUDE.md — built from actual failure points encountered during use, not theoretical warnings. The statusline customization tip (running a grammar-correction ad-hoc claude call inside the statusline script) demonstrates how Claude Code can be composed recursively for meta-tooling. The advice to use --setting-sources "" when calling claude from hooks prevents infinite recursion.

Connections

Raw Excerpt

The highest-signal content in your project CLAUDE.md (or any skill) is the Gotchas section. Build these from the failure points Claude Code actually runs into.

Graph View

Backlinks