本页以 FDE FAN 的第一视角讲解知识：先讲概念，再讲方法、图表、训练任务。底部保留参考来源，方便继续阅读。 This page teaches in FDE FAN's own voice: concepts first, then methods, visuals, and practice tasks. References are kept at the bottom for further reading.
知识结构地图Knowledge Structure Map
| 知识层级Knowledge Layer | 我们要掌握什么What We Need to Master | 训练产物Training Artifact |
|---|---|---|
| 核心论点Core thesis | AI 交付正在从单次项目进入可循环的生产飞轮，FDE 是把客户现场反馈送回产品和模型系统的人。AI delivery is moving from one-off projects into a repeatable production flywheel, with FDEs carrying field feedback back into product and model systems. | 把 FDE 训练成飞轮运营者，而不是一次性 Demo 执行者。Train FDEs as flywheel operators, not one-off demo executors. |
| 角色变化Role shift | AI 可以自动化部分配置、代码生成和文档生成，但现场语境、优先级判断、上线责任仍需要 FDE 承担。AI can automate parts of configuration, code generation, and documentation, but context, priority judgment, and launch accountability still need FDE ownership. | 训练学员判断哪些事情交给模型，哪些事情必须人工确认。Train learners to decide what the model can own and what requires human confirmation. |
| 组织意义Organizational meaning | 企业真正需要的是能把经验、评估、模板和交付机制沉淀下来的团队能力。Enterprises need team capability that preserves experience, evals, templates, and delivery mechanisms. | 每次项目必须输出复用资产。Every project must leave reusable assets. |
| 风险提醒Risk warning | 如果只追求模型能力或 Demo 速度，飞轮会断在上线、监控和复盘环节。If teams chase model capability or demo speed alone, the loop breaks at launch, monitoring, and retrospective. | 评分标准要覆盖日志、测试、回滚和复盘。Assessment must cover logs, tests, rollback, and review. |
核心观点Key Takeaways
FDE 的价值来自反馈速度：越靠近真实业务，越能发现模型和产品之间的缺口。FDE value comes from feedback velocity: the closer to real work, the faster product-model gaps appear.
AI 让部分部署工作自动化，但不会消除现场语境、边界判断和上线责任。AI automates parts of deployment, but it does not remove context judgment, boundary setting, and launch accountability.
我们训练“飞轮意识”：每次交付都要沉淀 Prompt、测试集、日志、模板和复盘。We train flywheel thinking: every delivery must leave prompts, evals, logs, templates, and retrospectives behind.
FDE 飞轮拆解FDE Flywheel Breakdown
| 层级Layer | 含义Meaning | FDE 产物FDE Artifact |
|---|---|---|
| 输入Input | 客户现场、数据、约束、成功标准Customer context, data, constraints, success criteria | 访谈表、资料清单、验收口径Interview brief, input inventory, acceptance definition |
| 编排Orchestration | Prompt、工具、Agent 状态、人工确认点Prompts, tools, agent state, human checkpoints | 工作流图、测试路径、失败回退Workflow map, test paths, fallback plan |
| 反馈Feedback | 日志、用户反应、错误样本、上线复盘Logs, user response, failure samples, retrospectives | 改版任务、评估集、交付文档Revision tasks, eval set, delivery docs |
深度讲义Deep Study Notes
1. 这不是“FDE 被 AI 替代”，而是 FDE 的工作对象变了1. This is not AI replacing FDEs; it changes what FDEs operate
“FDE-optional”容易被误读成企业不再需要 FDE。我们更准确地理解为，AI 让许多原来由 FDE 手工完成的部署动作变成可自动化组件：生成代码、整理文档、配置工作流、解释日志、批量改版。但当系统进入真实业务环境，仍然需要有人判断上下文是否完整、边界是否清楚、客户是否真的接受、结果是否能被业务验收。FDE 的工作对象从“亲手做每一个部署动作”，升级为“设计一个能不断吸收反馈的交付系统”。The phrase 'FDE-optional' can sound like companies no longer need FDEs. A better reading is that AI turns many formerly manual deployment activities into automatable components: code generation, documentation, workflow configuration, log explanation, and bulk revision. But once a system reaches a real business environment, someone still has to judge whether the context is complete, the boundaries are clear, the customer accepts the result, and the outcome can pass business validation. The FDE moves from doing every deployment step by hand to designing a delivery system that continuously absorbs feedback.
2. 飞轮的关键不是速度，而是反馈能不能回流2. The flywheel is not only about speed; it is about feedback returning to the system
我们要把每个项目拆成六个回流点：需求澄清回流到模板，Prompt 失败样本回流到提示词库，用户操作问题回流到前端交互，Agent 失败路径回流到工具编排，部署事故回流到运维清单，客户复盘回流到下一版项目任务。只有这些回流点存在，AI 才不是一次性加速器，而是能持续提高交付质量的飞轮。We should break every project into six feedback return points: requirement clarification back to templates, prompt failures back to the prompt library, user-operation issues back to UI interaction, agent failure paths back to orchestration, deployment incidents back to the operations checklist, and customer retrospectives back to the next project task. Only with these return points does AI become more than a one-time accelerator; it becomes a flywheel that improves delivery quality.
3. 面向学习者：FDE 是生产型角色，不是工具体验官3. For learners: FDE is a production role, not a tool tourist
我们要明确地区分 FDE 与普通 AI 工具课。普通课程教“如何使用工具”；FDE 训练必须教“如何把工具放进真实生产流程”。因此首页和课程页应强调部署链接、测试集、README、演示脚本、验收清单和复盘报告，而不仅是炫酷界面或 Prompt 技巧。We distinguish FDE training from ordinary AI tool classes. A normal course teaches how to use tools; FDE training teaches how to place tools inside real production workflows. The homepage and curriculum should therefore emphasize deployment URLs, eval sets, README files, demo scripts, acceptance checklists, and retrospectives rather than only polished UI or prompt tricks.
4. 企业诊断 Agent 的训练要求4. Training requirement for the Diagnosis Agent
诊断 Agent 不能只生成一份报告，还要把访谈、画像、报告、后续任务和复盘连接起来。用户每说一句话，都应该进入画像和机会地图；每次报告生成，都应该留下可追踪字段；每次建议，都应该能转成项目任务和验收条件。这样诊断工具本身就是一个 FDE 飞轮案例。The Diagnosis Agent should not merely generate a report; it should connect interview, profile, report, follow-up tasks, and retrospective. Every user statement should feed the profile and opportunity map; every generated report should leave traceable fields; every recommendation should be convertible into project tasks and acceptance criteria. In that sense, the diagnosis tool itself becomes an FDE flywheel case.
本课训练目标Training Goal
我们要把每个项目都训练成一个小飞轮：需求澄清、Demo、部署、观测、复盘、再交付，而不是一次性页面作业。We turn every project into a small flywheel: clarify, demo, deploy, observe, review, and ship again rather than submit a one-off page.
课堂任务Classroom Assignments
把一个已完成的项目拆成六个回流点，并写出每个回流点的可复用资产。Break a completed project into six feedback return points and write the reusable asset produced by each point.
为企业诊断 Agent 增加一张“飞轮日志表”：记录输入、模型判断、人工确认、输出和复盘。Add a flywheel log table to the Diagnosis Agent: input, model judgment, human confirmation, output, retrospective.
用 README 写清楚下一个 FDE 接手时应如何继续迭代。Use the README to explain how the next FDE should continue iteration.
引用Citation
HFS Research. The FDE-optional AI flywheel starts to spin. https://www.hfsresearch.com/research/fde-optional-ai-flywheel-spin/. Accessed 2026-06-28.