Ontology-Driven Design:如果本体成为 AI 生成代码的安全网?
当 AI 代理编写代码时,测试能通过,业务却可能悄悄漂移:同一个概念出现两个名字,某条规则被违背,一个已被弃用的观念死灰复燃。从代码中提取本体,在每次开发前先与意图比对,再与产出的代码比对,就能让这些漂移显形,并把有意的演进与错误区分开来。
当 AI 代理编写代码时,测试能通过,业务却可能悄悄漂移:同一个概念出现两个名字,某条规则被违背,一个已被弃用的观念死灰复燃。从代码中提取本体,在每次开发前先与意图比对,再与产出的代码比对,就能让这些漂移显形,并把有意的演进与错误区分开来。
每次开会都要花二十分钟才能让大家回到同一起点,有人一走交接就丢三落四:一个共同跟进的案子,靠重读邮件是接不回来的。一张在工作过程中随时维护的接手页——进展到哪里、哪些地方不牢靠、只写一个下一个行动——能让每个人,也能让 AI,在一分钟内从同一个位置重新出发。
“在简化的同时提高数据可靠性”这类综述能让所有人点头,却让真正的产品问题消失了。把每条客户反馈原样保存,一张卡片只写一个观点,再把张力本身写成一个独立的观点,一个令人尴尬的矛盾就变成了一个设计问题。原本天然倾向于抹平的 AI,也由此成为防止抹平的一道护栏。
一千条笔记彼此链接,却看不到全貌:哪些想法是核心,哪些是孤立的,两个相距甚远的概念如何相遇,都无从得知。如果 wiki 里每条链接的含义早已写在笔记中,一个 160 行的 Python 脚本就足以从中生成完整的图谱,再交给 Graphify 探索。整个过程只需 0.2 秒,不花一个 token。
你和 AI 的对话只留下结论,丢掉了理由:你排除过什么,哪条异议还悬而未决,你当初为什么那样决定。这本新书讲的,就是为解决这个问题而搭建的空间——一个主题一个空间,AI 能在里面找到一个想法的出处,并让工作从上次停下的地方继续。
关键在于知道下一步该检索什么,更在于知道何时停下。与其事先定下 30 或 50 个来源,不如从一个小而精选的语料库出发,从中提炼出结构化知识,再让它已经知道的和尚不知道的东西决定下一次检索。结果是:冗余更少,盲区可见,停止标准也可以检验。
一份文档本身什么都不产出,却约束着之后产出的一切,它该放在哪里?常用的标准——变化的快慢——不够用。取而代之的是一个按作用划分的标准:这份文件是在产出结果、保存知识、描述当前状态,还是在约束未来的行动?这个标准适用于架构未曾预想的任何对象。
花了几个月搭建的方法,会不会早已以别的名字存在于别处?查证的做法可以归结为几步:把想法拆解成机制,检索功能而不是词汇,走出本学科和英语世界,并事先规定一项新发现必须改变什么才算数。每一轮检索,可以宣称为原创的部分都在缩小,而方法本身的定位反而更清楚了。
不必依赖某处一台开着的 Mac,也能用 Contexte 工作。整套方案只有三块:ChatGPT 负责推理,项目指令定义引擎,Google Sheets 负责记忆。实际效果是:在展会上边走边口述一段对话,持久化,摄入,然后在离场前问系统还有哪些问题没问。
Five days later, you rarely find the one paragraph that mattered in a long ChatGPT conversation. This is the story of how I got a precise fragment out of the flow on demand, using nothing but a custom instruction and a Google Sheets outbox — a blocked API, a workaround that failed, then a solution that went from 50 seconds to two, with a systematic check after every write.
You can feel that a piece of writing or a decision isn't right yet, without being able to say why. That's where AI earns its place: it turns a vague unease into concrete handles — a blind spot, a shaky assumption, an objection you never dealt with. Which questions get you pushback rather than output, and why this makes human judgement more important, not less.
Every new format feels like starting over: you reopen the last document, copy, rephrase, and two versions slowly drift apart. The cause isn't a missing tool, it's treating a deliverable as your source of truth. How to install a reusable layer between your sources and your documents, and what AI gains from it.
Your best article ideas die between the moment they turn up and the moment you sit down to write. Here is the setup that catches them while they are still warm: the voice to empty your head without putting anything in order, a session source to keep the raw material raw, AI to surface the gaps in ten seconds rather than in two days.
Your team rebuilds the same product model for every question, every bug, every tender. Here is how to write it down once: several dozen objects, a few hundred rules, under six hours. And above all how to know you have finished — the three stopping criteria that look like proof of completeness and prove nothing, plus the one that holds.
AI raises the capacity to build, not the capacity to decide. Developers move up towards product; PMs move down towards building. The shift blurs roles without dissolving responsibilities. A full pass over what AI makes ordinary in product management, what stays hard, and a catalogue of emerging roles with the real risk attached to each.
Four days building a real product with Claude Code — 180 commits, starting from nothing. What it shows: the speed that goes to your head, the traps the AI creates on its own, what tests change when you haven't written the code yourself, and how far a PM can go without stepping outside their remit. A field report, not a doctrine.
Producing code faster is worth nothing if nobody understands what got built. AI automates precisely the tasks that used to train junior developers, and an error that lands too late no longer teaches anything once the system has grown too big to analyse. A few concrete anchors for staying in control: clear contracts between APIs, tests tied to the functional need, rebuildability as a quality bar rather than raw delivery speed.
Your documents never say everything. The fine-grained understanding of what happens on the ground stays in the PM's head. Session sources open a channel for pouring that informal knowledge into an AI Context — a hallway remark, a nuance picked up from a customer, a tension just starting to show — and turn dozens of micro-signals into notes, assumptions and better-framed decisions.
Your tools file information away neatly, but they never build up what that information teaches you. This piece explains why the future of the information system is not a better AI-assisted form, but a shift towards `Contexts`: portable, governed units of memory that connect sources, decisions and reasoning instead of merely storing them.
Do you lose half an hour in a test environment just to answer a product question on Slack? That hidden cost disappears once AI can query the source code, the real truth about how the product behaves. Answers grounded in evidence — functions, tests, commits — rather than in documentation that has fallen behind. Provided the code speaks the language of the business.
Do you lose the thread the moment you reopen a conversation with an AI about something that has been running for weeks? This article lifts the lid on a structured memory system: the difference between the goal and the attention of the moment, the directives, the offline queue, the three memories (log, synthesis, checkpoints) and the cycle of a session. Enough to see how an AI can genuinely remember a mission.
Most organisations know how to produce documents. They don't know how to bank the reasoning behind them. On every decision they start again almost from scratch. The PM as Context architect turns scattered customer knowledge — transcripts, objections, past calls, weak signals — into a reusable asset. The result: faster, sturdier decisions, and an advantage competitors cannot copy overnight.
Classic RAG starts from scratch with every question. By building an AI Wiki — a persistent wiki maintained by an LLM — you capitalize on domain knowledge once and for all. Atomic notes, links between concepts, contradiction detection: knowledge is compiled, not rediscovered. Lessons learned from 950 notes extracted from industrial maintenance sources, with calibration insights and concrete use cases.
Drowning in email and losing track of the threads in flight? By handing the structuring over to Claude — one file, a few directives, then a gradual split — I turned 500 building-management emails into a queryable knowledge base. Seven steps, from the vague directive to a multi-file system, without ever needing a complex framework.
Documentation never keeps up with the code. We know it, we accept it, and yet it costs us time with every change. AI shifts the equation: when the code is clean, it becomes the source of truth from which the other artifacts—specs, changelogs, support documentation—can be regenerated. This code-centric approach demands a minimum of technical literacy, but it opens a real lever for the product managers who master it.
A wasteful skill locks you out after 3-4 runs instead of 8-10. Discover the 5 silent traps that saturate your quota, how to audit in 3 simple phases, and how to gain 40% more capacity without changing your model. Productivity regained.
Producing software is no longer rare. The cloud commoditized it, and AI will accelerate the trend further. When everyone can build fast and well, churning out features by the yard is no longer enough. The real competitive edge shifts toward understanding the context: the business, the users, the customers, the competition, the standards. Understanding better than others, before building.
Your days as a PM are devoured by tasks that have nothing strategic about them. In eight days on Claude Code, I automated my competitive intelligence — gathering, organizing, updating. Time reclaimed to analyze, understand, decide. This field report shows concretely why the center of gravity of the Product Manager job is shifting: fewer intermediate artifacts, more real strategy.
No more re-explaining the context at every session. Claude automatically retains what you did, the decisions made, the problems encountered — per project, inside your git repo, with no database. Portable, versionable, under $1/month. Here's how to set this system up in less than an hour.