/spec先于代码规则显著提升代理代码质量并减少调试时间，本地LLM如Qwen3 32B实测有效
英文摘要
A Reddit user tested agent-skills frameworks such as addyosmani/agent-skills and obra/superpowers with a local Qwen3 32B model. The key insight is that forcing the agent to write a specification before any code catches design flaws within two minutes—rather than spending two hours debugging—and substantially raises code quality by avoiding guesswork. The /plan - /build - /test pipeline keeps each step bounded, which suits local LLM workflows well, and overall token usage drops because the agent no longer generates multiple incorrect implementations before arriving at the correct one.
中文摘要
一位Reddit用户使用本地Qwen3 32B模型测试了addyosmani/agent-skills和obra/superpowers等代理技能框架。关键发现是强制代理先编写规范说明可在两分钟内发现设计缺陷，而非花两小时调试，同时显著提升代码质量，避免猜测式编程。/plan - /build - /test流程使每个步骤上下文边界清晰，适合本地LLM工作流，且总体token消耗减少，因为代理不再产生多个错误实现后才得到正确版本。
关键要点
The /spec before code rule catches design flaws in minutes and improves code quality by preventing the agent from guessing what the user wants.
/spec先于代码规则能在几分钟内捕获设计缺陷，通过防止代理猜测用户意图来提升代码质量。
Using Qwen3 32B locally, the workflow reduces overall token usage by eliminating multiple wrong implementations before the correct one.
在本地使用Qwen3 32B时，该工作流通过消除得出正确实现之前的多次错误尝试，减少了总token消耗。
Agent-skills frameworks like addyosmani/agent-skills and obra/superpowers implement a bounded-context pipeline (/plan, /build, /test) that works well with local model constraints.
agent-skills等框架实现了上下文边界清晰的/plan、/build、/test流水线，适合本地模型的限制条件。