This are some of my personal recommendations. They work for me, but it doesn't mean it will work for everyone. And its constantly evolving list like any other AI tools.

• Use the superpower plugin.
• Remove MCPs that aren’t needed.
• Use a prompt/workflow like: “clarification → plan → approve → implement.”
• Keep a global CLAUDE.md at ~/.claude/CLAUDE.md that’s generic for all projects.
• Keep a local CLAUDE.md for project-specific rules.
• Have code written by Claude Code (Opus 4.5), and reviewed by Codex (GPT-5.2).
• Review extensively—spend half as much time (or the same amount of time) reviewing as you do coding.
• Make small commits; don’t run multiple Claude Code sessions into the same commit.
• Use only the MCPs you need (e.g., for me: HF, Zotero). Don’t use unnecessary ones, since they eat up system prompt.
• Add .claude to your dotfiles (and if you dont use dotfiles, start doing so)

References

1. https://x.com/affaanmustafa/status/2012378465664745795
2. https://github.com/affaan-m/everything-claude-code
3. https://github.com/forrestchang/andrej-karpathy-skills/blob/main/CLAUDE.md
4. https://www.anthropic.com/engineering/claude-code-best-practices
5. https://anthropic.skilljar.com/claude-code-in-action?next=%2Fclaude-code-in-action%2F303235