The Top Open‑Source Companion for Claude Code & Codex in IDEA (3.3k★)

Repository overview

The GitHub repository https://github.com/shanraisshan/claude-code-best-practice (over 54 k stars in less than two months) organizes community‑validated Claude Code best practices into four hierarchical layers: concepts → features → workflows → tips.

Difference from official documentation

Official docs describe command semantics (e.g., /compact). The repository adds empirical guidance: Claude’s comprehension degrades after roughly 300 k–400 k tokens, and total context utilization should stay below 40 %. When the status line shows utilization above this threshold, run /compact immediately.

Four main sections

Core concepts

Ten key concepts are covered: Subagents, Commands, Skills, Workflows, Hooks, MCP Servers, Plugins, Settings, Status Line, and Memory. Each concept links to concrete examples. For MCP Servers, the most useful MCPs listed are chrome-devtools-mcp (browser debugging), github-mcp (PR operations), and sequential-thinking (forced step‑by‑step reasoning).

Popular features

The section tracks the latest Anthropic beta capabilities, including Ultrareview, Devcontainers, Channels, Ultraplan, Auto Mode, Computer Use, Agent SDK, Code Review, GitHub Actions, Agent Teams, and Scheduled Tasks. It provides early‑adopter experiences before official documentation is updated.

Development workflows

Thirteen well‑known Claude Code workflows (e.g., Superpowers, Everything Claude Code, Spec Kit, gstack, Get Shit Done) are compared in a table that records star count, uniqueness, presence of a planning stage, and the number of Subagents/Commands/Skills used. The analysis identifies a common pattern across successful workflows: Research → Plan → Execute → Review → Ship.

Practical tips

Eighty‑three tips are organized into fifteen categories (Prompting, Planning, Context, Session Management, Memory, Agents, Commands, Skills, Hooks, Workflows, Advanced Workflows, Git/PR, Debugging, Utilities, Daily practices). Three especially actionable tips are:

Keep context utilization under 40 %; when token count reaches ~300‑400 k, run /compact immediately.

Separate Plan and Execute phases. After planning, commit the plan to a markdown file, start a new session, and have Claude read that file before execution to ensure a clean context.

Use Hooks to enforce discipline rather than add features. A PostToolUse hook that runs mvn test whenever Claude modifies files under src/ prevents the common “forgot to run tests before commit” pitfall in multi‑module Spring Boot projects.

A cross‑model combination is also described: Claude Code generates code while Codex performs code review, providing a more objective assessment than a single model.

Security and transparency FAQ

Recent controversy stemmed from hidden marker information inserted into system prompts, which breaches user trust. The issue is a transparency problem rather than a raw capability problem; users of high‑privilege agents care about clear disclosure of agent behavior. Anthropic team members indicated the problematic logic will be removed in the next version, but long‑term trust requires demonstrable transparency and controllability.