李彦卿的个人技术站，记录人工智能质量工程、智能体治理、AI 原生测试自动化与可验证的软件交付实践。https://liyanqing.xflux.cn/https://liyanqing.xflux.cn/articles/shared-work-state-ai-collaboration/https://liyanqing.xflux.cn/articles/shared-work-state-ai-collaboration/AI 协作真正缺少的，往往不是更强的模型，而是一份不会随着聊天结束而消失、能被人和不同 AI 持续维护的当前工作状态。Sun, 19 Jul 2026 00:00:00 GMThttps://liyanqing.xflux.cn/articles/loop-engineering/https://liyanqing.xflux.cn/articles/loop-engineering/Loop Engineering 关注的不是把提示词写得更长，而是设计跨多轮运行的控制结构：任务如何触发，状态如何保存，结果如何验证，以及系统何时重试、停止或转交给人。Sat, 20 Jun 2026 00:00:00 GMThttps://liyanqing.xflux.cn/articles/shift-left-ai-product-quality/https://liyanqing.xflux.cn/articles/shift-left-ai-product-quality/AI 产品团队做测试左移，重点是从需求阶段定义质量标准，并提前确认 prompt、数据、模型输出、风控规则和回归方式。Thu, 18 Jun 2026 00:00:00 GMThttps://liyanqing.xflux.cn/articles/llm-output-assertion/https://liyanqing.xflux.cn/articles/llm-output-assertion/LLM 输出天然存在非确定性，测试不能只依赖全文匹配。更可靠的策略是组合结构化断言、关键事实校验、禁止项和语义相似度。Wed, 17 Jun 2026 00:00:00 GMThttps://liyanqing.xflux.cn/articles/ai-test-framework-playwright-pytest-allure/https://liyanqing.xflux.cn/articles/ai-test-framework-playwright-pytest-allure/AI 产品的自动化测试不只是记录页面操作，还要记录输入、断言、失败证据和报告，方便后续复查。Tue, 16 Jun 2026 00:00:00 GMThttps://liyanqing.xflux.cn/articles/harness-engineering/https://liyanqing.xflux.cn/articles/harness-engineering/从 Prompt、Context 到工具权限、验证门禁和反馈回路，解释 Harness Engineering 如何把 AI coding agent 纳入可控的软件工程流程。Tue, 16 Jun 2026 00:00:00 GMT