使用裁判子代理和多阶段审查构建应用程序
Build apps with a judge subagent and multi-stage review
使用 Claude Code 完成。一个用于应用程序开发的 Claude Code 设置使用裁判子代理、上下文管理和多阶段审查进行质量控制
做了什么
The post describes a solo developer using Claude Code to build a Windows app with a separate read-only judge subagent for unresolved decisions. The judge had to label claims as confirmed, guessed, or unknown and avoid deciding from guesses; the developer also started each issue in a fresh child session, then ran two review passes on code: a general code review and a second pass using coding-guideline skills, with deferred issues left in PRs for human review.
效果如何
The workflow reportedly led to a Microsoft Store shipment and 142 merged PRs in 17 days. The post also notes the setup is complex and may be hard for beginners.
值得注意
The main catch is overhead: this workflow adds fresh sessions, a separate judge, and two review stages, so it is presented as an advanced setup rather than a lightweight default.
试试这个 Prompt
Set up my project in [repo path] so that after you implement each feature of [app description], a separate judge subagent reviews the work against the requirements and tests and sends back a list of problems. Repeat the fix and review loop until the judge has no blocking issues. Keep a short notes file of decisions so context survives between sessions, and summarize the final state when finished.