An Embarrassing Situation from a Manager's Perspective

We're a small team of seven. Engineers document technical designs in Notion, the product manager writes requirements in Feishu docs, I'm in the habit of saving links in WeChat Favorites, and someone else writes meeting notes in local Markdown files.

This "free-for-all" approach didn't seem like a problem—until a new hire joined and I asked him to "take a look at our past lessons learned." He asked me back: where exactly?

That's when it hit me: the team's knowledge assets existed, but they weren't structured—so effectively they didn't exist. The bigger issue was the risk dimension—the context behind a key decision was scattered across five discussion threads, two documents, and a chat log. When it came time for a retrospective, nobody could piece together the full chain. If knowledge can't be traced, accountability becomes impossible to establish.

A manager's pain points boil down to three things:

1. Scattered information: multiple tools, multiple formats, multiple personal habits, no unified entry point;
2. Low reuse rate: the same problems get hit two or three times by different people, because nobody can dig through that pile of raw notes;
3. Uncontrolled risk: decision rationale, lessons learned, commitments to clients—"organizational memory" stored in personal accounts leaves with the person.

I tried to force a unified note-taking standard. It failed within two weeks—the cost of following the standard fell on everyone, and nobody wants to pay "extra hours now" for "possible retrieval later."

Then I handed this over to AI, and my thinking changed: don't ask people to change how they take notes—just add an AI processing step downstream.

My Workflow: Three Steps to Consolidate

Step One: Collection, Without Touching Existing Habits

Every Friday, each person exports their notes for the week (Notion to Markdown, Feishu to Word, chat logs copied and pasted) into a shared folder. This step only requires one action—"export"—with no format requirements and no template requirements, so friction is minimal.

Step Two: AI Structuring—This Is the Core

I feed all the raw material into a model that supports long context, using a fixed prompt to turn it into "knowledge cards." The card fields were locked in after team discussion, so the cards are naturally searchable and archivable:

你是一名团队知识管理助手。请将下面提供的零散笔记整理成结构化知识卡片。

【任务要求】
1. 从笔记中识别出独立的"知识单元"（一个决策、一个踩坑、一个方案、一个结论均算一个单元）。
2. 每张卡片包含以下字段：
   - 标题：一句话概括，不超过25字
   - 类型：决策 / 踩坑 / 方案 / 待确认
   - 背景：为什么会有这条记录，1-2句
   - 核心内容：关键事实与结论，3-5条要点
   - 关联：与其他卡片可能的关系（如有）
   - 信息来源：来自谁的笔记、日期
   - 置信度：高（多人记录或已验证）/ 中（单人记录）/ 低（存疑，需复核）
3. 无法确认的信息不要自行补全，标记为"待确认"类型。
4. 涉及数字、日期、人名的信息，逐字保留原文，不要改写。

【输出格式】
以 Markdown 输出所有卡片，每张卡片用 --- 分隔。

【原始笔记】
（在此粘贴内容）

Two details are worth mentioning. First, the "confidence" field—this is key to risk control: content recorded by a single person and unverified gets explicitly marked "low confidence" on the card, reminding readers to double-check and preventing personal recollection from being treated as team fact. Second, "no autonomous fill-in": models are very good at "hallucinating plausible details," but in a team knowledge base, that's poison.

Step Three: Manual Spot-Checks and Archiving

I don't proofread word by word (that would defeat the purpose), but I do two things: spot-check the low-confidence cards, and confirm whether the "pending confirmation" items get supplemented by the responsible leads. After checking, everything goes into the team wiki's weekly directory. The whole process takes about 40 minutes per week—30 minutes of AI processing and only 10 minutes of human time.

Before and After AI

What exceeded expectations most was the change at the collaboration level: because "lessons learned" cards are visible to everyone, the second person to run into the same problem often searches the card library first. Repeated mistakes dropped noticeably—something completely impossible in the era of raw notes, because nobody reads anyone else's stream-of-consciousness notes, but everyone is willing to read a card that takes 30 seconds.

Three Management-Level Lessons

First, AI is a processing layer, not a recording layer. Don't try to use AI to change how the team takes notes—just let it turn "existing chaos" into "usable order." In our team, any solution that required members to change their habits didn't survive a month.

Second, confidence labeling is the bare minimum for risk control. Once team knowledge is structured, it gets cited as fact. Having AI clearly distinguish "verified" from "questionable" costs far less than correcting errors after the fact. Anything involving contract figures or commitment dates must be preserved verbatim.

Third, the process needs a fixed rhythm. Those 40 minutes every Friday are a fixed item on my calendar. With knowledge consolidation, once you skip two weeks, the backlog piles up to the point where nobody wants to catch up.

On cost: if you run this process through API calls, long-context processing consumes a significant number of tokens. Check the official pricing page for actual costs—I recommend running one week's worth of real notes as a test before deciding whether to process everything or only key projects.

Final Thoughts

For a manager, the essence of this isn't "using an AI tool"—it's turning the team's unordered individual memories into a risk-controlled organizational asset at extremely low process cost. If you want to build your own AI processing pipeline, you can start by registering here: https://api.thistoken.ai/register

---

Every example in this post runs with a single API key — get yours at https://api.thistoken.ai/register and start in minutes.