AI × 敏捷
作品与输出 AI × Agile
works & writing

《AI Coding 需求全链路行为效能挖掘》

传统效能指标在 AI 人机协同下全面失效——放弃结果考核，用 DEED 模型做全链路『无痕』行为流动观察。Legacy efficiency metrics break down under AI pair-work — DEED replaces outcome policing with frictionless observation of behaviour flow across the whole chain.

• 产出 / 工作量 / 过程 / 吞吐四类旧指标，在 AI 时代全部失效Four families of legacy metrics all fail in the AI era
• DEED = 行为驱动 + 流动驱动 + 价值驱动，只看真实操作行为DEED: behaviour-, flow- and value-driven — real actions only
• 2026 迭代：AI 升级为自主 Agent，新增行为偏差识别与人机断点挖掘2026: agents act autonomously — add deviation detection & human-machine break analysis

「效能观察，不看『快不快』，只看『顺不顺、稳不稳』。」

《产研团队 AI Coding 落地与项目经理实操》

AI Coding 不是“更快写代码”的工具采购，而是一次以高质量上下文、可验证规范与自动化质量能力为中心的交付系统重构。AI Coding is not a tool purchase for “faster code” — it is a rebuild of the delivery system around high-quality context, verifiable specs and automated quality.

• 探索与生产分轨：Vibe 低成本试错，Spec 作为可验证契约做生产交付Split exploration from production — Vibe to test cheaply, Spec as a verifiable contract for delivery
• 上下文是团队资产：治理“最小可行上下文包”，而非只培训 PromptContext is a team asset — govern a minimal viable context pack, not just Prompt training
• 质量能力是 AI 产能的一部分：提速前先建门禁与自动化，避免测试堰塞湖Quality is part of AI capacity — build gates and automation before speeding up

「AI 更像组织能力的放大器——成熟平台与清晰流程才把局部提效变成业务结果。」

《AI Coding 越改越乱？开发 Skill 的 2 天，踩了 4 个大坑》

一次真实的 Skill 开发复盘：失控的根因不是模型不行，而是缺了它周围的 harness。A raw post-mortem of 9 rework rounds: loss of control isn't the model — it's the missing harness around it.

• 四个坑是同一种病：整条链路唯一的质检员是你的眼睛Four pits, one disease: your eyes are the only QA in the loop
• 规则写进文档只是『建议』，写进校验脚本才是『强制』Rules in docs are advice; rules in scripts are law
• 可靠性三阶梯 L1 建议 / L2 强制 / L3 验证闭环，每升一层可控性上一个量级L1 advice / L2 enforcement / L3 verification — each step orders of magnitude safer

"Memory is advice. Hooks are law." — Mitchell Hashimoto

《浅谈敏捷团队度量指标分析》

度量之所以必要，是为了『显性』与『统一』；而分析的要义，是跨指标的系统性解读。We measure to make work visible and language shared; the craft lies in reading metrics as a system.

• 指标相互制衡：拉满任何一个，总免不了拖累别处Metrics constrain each other — max out one, pay elsewhere
• 生产力 / 质量 / 稳定性三类指标环环相扣，须联动解读Productivity, quality, stability — read them as interlocking loops
• 度量重点是看偏差，不是绝对好坏；北极星指标也有陷阱Measure deviation from goals, not absolutes — even north stars mislead

「一个指标会说谎，一群指标会告诉你真相。」

《从蛰伏到狂飙的 Scrum Master》

国内 SM 几乎全为兼职的现实下，把 SM 从『沉没成本』变成『撬动组织效能的杠杆』。In a world of near-100% part-time Scrum Masters, turn the role from sunk cost into an organisational lever.

• 调研：99% 公司的 SM 为兼职，定位分 As a SM / PM / Aid 三种Survey: 99% part-time SMs, positioned as SM / PM / aide
• 众生相三型：小鸟型引导、老牛型服务、狮子型真正带队Three archetypes: bird facilitator, ox servant, lion leader
• 好的兼职 SM，是组织储备研发管理人才的资源池A good part-time SM is the talent pool for engineering management

「敏捷灵魂之一，是对人的关注，以人为核心驱动。」

《有模有样就是高绩效团队？》

记一次『激活』与『交融』的团队教练工作坊设计：周年庆不是仪式，是精心设计的场域。An anniversary as coaching design: not a ceremony, but a deliberately built field for activation and fusion.

• 准备四步：摸底（one-one 访谈）· 策略 · 勾兑干系人 · 源泉备素材Prepare in four: interviews, strategy, stakeholder sync, source material
• 场域设计：议题构建『高→低→高』情绪曲线，数据与讨论穿插抛出钩子Design the field: a high-low-high emotional arc with data woven into talk
• 提问从易到难三段递进，并备好冷场 Plan BQuestions escalate from easy to soul-probing — with a Plan B

「毕竟，真正的高绩效团队，是需要『团魂』的。」

《『平湖起涟漪』的团队工作坊设计灵感》

如何打破团队的表面平静而不引发冲突：两个观念 + 四个工具。Breaking a team's surface calm without triggering conflict: two mindsets, four tools.

• 工作坊是一场『真心话大冒险』：安全场域 + 不可控的走向A workshop is truth-or-dare: safe field, uncontrollable outcomes
• 摸着石头过河：流程不可照搬，『石头』就是团队状态Feel the stones: never copy agendas — the stones are the team's state
• 四工具：破冰卸防 · 创造对话 · 回忆唤醒 · 冲突共情Four tools: ice-breaking, dialogue, memory, empathy

「要摸的『石头』就是团队的『底』，就是团队状态。」

PMA · 项目交付助手

轻量项目排期与交付记录 Agent：需求变更先预演、确认后入排期，事件留痕、周期自动计算，不引入复杂项目管理系统。A lightweight scheduling & delivery-log agent: preview changes before committing, log milestones, auto-compute cycles — no heavyweight PM system.

• 需求变更预演：自然语言描述变更，预估排期影响，不修改正式计划Change preview: describe in plain words, see schedule impact before committing
• 项目事件留痕 + 交付周期自动汇总，形成完整项目历史线Milestone logging with auto-computed delivery cycles — a full project history
• CSV 本地留存：可查看、备份、迁移、审计，规则可解释CSV-first local storage: viewable, portable, auditable, explainable

敏捷团队季度 Review 分析台Quarterly review workbench

为敏捷团队季度 Review 打造的本地比对分析台：上传两期 xlsx，等级分布、治理统计、走查对比一次生成。A local review workbench for agile teams: drop in two periods of xlsx, get distribution, governance and walkthrough comparisons at once.

• 六大模块：团队度量 / 成熟度分布 / 需求岛治理 / 需求池治理 / 走查对比 / 说明文档Six modules: metrics, maturity, governance ×2, walkthrough, docs
• 双期对比是默认能力：两期文件并列生成图表与差异洞察卡片Two-period comparison by default — side-by-side charts with insight cards
• xlsx 进、评审材料出：等级饼图、规范度对比、逐条明细自动生成xlsx in, review deck out — pies, compliance charts and drill-downs, all auto-built

Journal Generator · 照片转手帐

开源 Agent Skill：一组照片生成整页视觉手帐——Agent 理解图像撰写克制文案，Python 只做确定性排版。Open-source Agent Skill: a photo set becomes a polished visual journal — the agent writes captions, Python lays it out deterministically.

• 两种版式：S 形照片流（叙事向）/ 双栏清单卡（合集向）Two layouts: S-shaped flow / two-column checklist
• 文案语言自动跟随用户，手写 captions.json 完整保留Caption language follows you; hand-written captions.json preserved
• Codex / Claude Code 即装即用，排版渲染零外联Drop-in for Codex / Claude Code; rendering needs no network

Photo Zine Collage · 照片转 Zine 拼贴

开源 Codex Skill：照片转 3:5 竖版撕纸拼贴 Zine——照片是事实锚点，调色板取自源图，撕纸与油画笔触向外延展。Open-source Codex Skill: a photo becomes a 3:5 torn-paper zine collage — the photo is the factual anchor, the palette comes from the source.

• 两种视觉模式：印象派大色块 / 写实细线，调色板加权取自源图Impressionist blocks or realist fine lines — palette weighted from the source
• 八种源图版式 + 英文微文案确定性生成，可指定原文或免文字Eight layouts + deterministic micro-text, exact-wording or text-free
• 原图像素精确还原，签名与水印完整保留Source pixels precisely restored; signatures and watermarks kept