Teaching AI assistants to stop and think before they code
A Chinese-localized fork of the popular superpowers framework that makes 18 AI coding tools ask clarifying questions instead of immediately generating wrong code.

What it does
superpowers-zh is a localization and extension of the 159k-star superpowers project. It installs 20 “skills” — structured prompt templates — into AI coding tools like Claude Code, Cursor, Windsurf, and 15 others. The core pitch: instead of an AI immediately writing code when you say “add a batch export feature,” it first asks about format, data volume, and permissions, then proposes 2-3 approaches.
The interesting bit
The project isn’t just translation. It adds 6 China-specific skills covering Gitee/Coding/CNB Git workflows, Chinese commit conventions, and Chinese documentation typography rules. Four of these require manual invocation (/chinese-code-review) to avoid interfering with the upstream auto-scheduling system — a thoughtful bit of prompt engineering hygiene.
Key highlights
- One
npx superpowers-zhcommand auto-detects your project’s AI tools and installs to the correct directories (.cursor/skills/,.kiro/steering/, etc.) - 14 translated upstream skills (TDD, systematic debugging, subagent-driven development, Git worktrees)
- 2 auto-triggered original skills: MCP server builder and YAML workflow runner for multi-role orchestration
- Covers domestic tools often ignored by Western projects: Trae, Qwen Code (通义灵码), Qoder, DeerFlow 2.0
- Maintains sync with upstream while accepting PRs for Chinese developer pain points
Caveats
- Manual
cp -r skillsis explicitly labeled “low-fidelity” — skills exist but won’t auto-trigger without the full plugin’s hooks and bootstrap files - Running in
~is blocked since v1.2.1; older versions would pollute your home directory withCLAUDE.mdand skills - The 4 Chinese cultural skills are opt-in only; the AI won’t automatically use Chinese code review style or domestic Git platforms
Verdict
Worth a look if you’re using any of the 18 supported tools and find yourself constantly correcting AI-generated assumptions. The localization effort is genuine, not a thin translation layer. Skip it if you’re already happy with the upstream English version and don’t need Gitee workflows or Chinese commit conventions.