The Old Manager's Problem: Knowledge Is Slipping Away

Anyone who leads a team is probably familiar with this scenario.

After a Monday product review meeting, a team member takes notes and produces a meeting summary of over two thousand words—discussion process, decisions, and action items all mixed together. Three months later, when a new member joins and wants to understand "why we decided to do it this way back then," they search through the shared folder and can't find the answer. Only after asking the people involved do they learn: that decision was recorded in the third paragraph of some weekly meeting notes, sandwiched between two action items.

The notes weren't unrecorded—they were recorded in a way that's as good as not recording them. This is the most typical way knowledge is lost in small teams:

• Uneven information density: A single paragraph mixes background, discussion, conclusions, and action items; finding the key points requires eyeballing everything
• Structure varies from person to person: Everyone has different note-taking habits, and trying to unify everything later is a massive amount of work
• Unclear ownership: Who is responsible for organizing? Who is responsible for archiving? Without a process, no one does it
• Extremely low reuse rate: The same judgment was made last year, only to be re-discussed from scratch this year

My principle is this: organizing knowledge shouldn't take up engineers' productive time, but it also can't be skipped. So there's only one answer—hand it over to AI for process-driven handling, with humans doing only the final quality check.

II. The Workflow I Designed

I built it as a four-step pipeline, where AI does the heavy lifting at each step and humans make the decisions.

Step 1: Raw notes go into a queue. All meeting records and casually jotted fragment notes are dropped into a unified directory as-is, with no preprocessing whatsoever. This point matters for managers—don't make team members spend time "tidying things up first before handing them to the AI"; that's duplicated effort.

Step 2: AI-based structured extraction. Using a unified prompt template (see below), have the AI break each note into standard cards: fact cards, decision cards, action item cards, and risk cards. Each card includes a source citation that can be traced back to the specific paragraph in the original note.

Step 3: Manual spot-checking. I don't review everything line by line—I only spot-check 10%. I focus on two things: whether the AI has misclassified "speculation during discussion" as "confirmed decisions," and whether the responsible person and deadlines for action items were extracted incorrectly. These two types of errors have the biggest impact and must be caught by a human.

Step 4: Structured storage and regular retrospectives. Cards are stored with project tags, and once a month the AI produces a summary—a "decision list for this month" and "unclosed action items"—which serves as input for the monthly retrospective.

Once the whole workflow runs smoothly, the labor investment in the organizing step approaches zero, and people only spend time on spot-checking and retrospectives.

III. The Core Prompt Template

This is the template used in Step 2. You can copy it directly and replace the placeholders:

你是一名知识管理助手。请把下面的原始笔记整理成结构化知识卡片。

要求：
1. 将内容拆分为四类卡片：
   - 【事实卡片】客观信息：数据、背景、现状描述
   - 【决策卡片】已达成结论的事项，必须注明决策依据
   - 【待办卡片】行动项，格式为：负责人 / 截止时间 / 具体动作
   - 【风险卡片】讨论中提及但未解决的隐患或分歧
2. 每张卡片必须引用原文对应句子，标在「来源」字段
3. 分辨不清是"建议"还是"决定"的内容，归入风险卡片并注明存疑
4. 不新增原文没有的信息，不做主观推断
5. 输出为Markdown表格，字段：类型 / 内容 / 来源 / 备注

原始笔记：
{粘贴你的笔记内容}

Rules 3 and 4 are the key risk controls. The most common problem with AI isn't incomplete extraction—it's "over-imagining," writing up discussion tendencies as conclusions. These two constraints, combined with source citations, keep hallucination risk within a controllable range and also provide a basis for verification during spot-checks.

IV. Before and After Using AI

From a management perspective, the biggest change isn't saving time—it's the controllability of knowledge assets. Before, team knowledge lived in each person's memory and individual documentation habits; when someone left, the knowledge was lost. Now it's captured in a unified format that is searchable, traceable, and reviewable.

V. A Few Risk Control Reminders

• Don't let the AI directly write conclusive "meeting decisions." The AI only does extraction and classification; conclusions must be traceable to the original text. The moment a team starts trusting "decisions summarized by the AI," the risk begins.
• Stick with the spot-check mechanism. Even if AI accuracy is high, spot-checking matters because it keeps humans' judgment sharp and helps detect prompt failures early (e.g., degraded extraction quality after note formats change).
• Classify sensitive information. For notes involving compensation, personnel matters, or client privacy, first confirm the data handling policies of the AI service you use before deciding whether to send them in, or anonymize them first.

VI. Final Thoughts

This workflow itself isn't complicated—the hard part is choosing a stable, controllable AI service to run it. If your team is evaluating AI API platforms, take a look at Thistoken (https://api.thistoken.ai/register); unified API access makes it much cheaper for small teams to implement this kind of batch processing task. For specific models and pricing, refer to the official pricing page.

Hand the tedious work of organizing knowledge to an AI workflow, and let people focus on judgment and decision-making—this is one of the most practical ways for a small team to use AI.

---

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