The Meeting Is Over — Now the Tiring Part Begins

Independent developers and small teams have probably had a similar experience: after a one-hour product alignment meeting, you want to leave behind a usable record, so you start replaying the recording, fast-forwarding, rewinding, and taking notes. What's truly time-consuming isn't the meeting itself — it's the post-meeting write-up.

I once did a rough count myself: for a 60-minute recording, manually organizing "what was discussed, what was decided, who is responsible for what" takes an average of 90 to 120 minutes. And the quality of the write-up depends on your patience at the moment — by the second half, you often just jot down a few key points, and critical action items easily slip through the cracks.

What makes meeting notes hard isn't the "recording" — it's the "layering":

• Decisions hide inside discussions. An offhand "let's go with option A" gets forgotten by everyone once the meeting ends.
• Action items are scattered everywhere. "I'll look into that later," "let's sync up next week" — nobody writes them down when they're said, and nobody claims them afterward.
• Background noise dominates. Pleasantries, digressions, and repeated arguments take up more than half the recording but add zero value to execution.

Manually organizing notes is, in essence, doing exactly what machines are best at: extracting structured information from a large volume of low-density speech.

My Current Workflow: Recording In, Two Lists Out

The workflow now is very simple — four steps in total:

Step 1: Get the transcript of the recording. The meeting tool's built-in transcription, or a locally-run speech recognition model — either works. This step only needs to produce plain text; the quality doesn't have to be perfect, and a few typos won't affect the subsequent extraction.

Step 2: Feed it to the model and extract with a fixed template. This is the core step. The prompt template I use is at the end of this post and can be copied directly. It forces the output of two lists: a decision list (each decision with context and rationale) and an action item list (each item with owner, deadline, and dependencies).

Step 3: Spend five minutes on manual review. Focus on two things: whether the decision list contains any "over-generalized" items (mistaking discussion for conclusions), and whether the owners in the action item list are correct. This step can't be skipped, but the workload is tiny.

Step 4: Distribute. Archive the decision list to the project documentation, and push the action item list directly into the task management tool.

Before vs. After: Time Cost Dropped by an Order of Magnitude

Taking a 60-minute meeting as an example, here's my own before-and-after data:

At three meetings per week, that saves roughly 12 to 15 hours of organizing time per month — nearly two full workdays. For a team of three to five people, this is almost like gaining an extra person's meeting-support capacity out of thin air.

Beyond time, there's a hidden benefit: standardization of records. Meeting notes used to vary wildly in format depending on who wrote them, making digging through history painful. Now every meeting's output format is identical, and "what did we actually decide in that meeting three weeks ago" becomes a ten-second lookup.

What the Model Does Better Than Humans

• Full scan, no laziness. Humans get fatigued halfway through organizing; models don't. A decision that flashed by verbally gets caught, as long as it's in the transcript.
• Distinguishing "discussion" from "decision." A good prompt makes the model extract only definitive conclusive statements, compressing the back-and-forth arguments into a one-line background note.
• Filling in the to-do structure. Turning vague statements like "I'll look into it later" into "who, does what, roughly when" based on context — precisely the part task management needs most and where humans are most prone to cutting corners.
• Potential for cross-meeting correlation. Once decision lists accumulate, you can have the model do another layer of aggregation: which decisions keep getting overturned and re-debated, which to-dos have gone unclaimed for three weeks. This is nearly impossible to sustain with manual note-taking.

A Reusable Prompt Template

你是一名会议记录整理助手。我会给你一份会议录音的转写文本，请严格按照以下要求输出：

一、先通读全文，区分三类内容：明确的决策、待办事项、讨论过程。

二、输出「决策清单」：
- 每条决策一行，格式：【决策】+ 一句话结论
- 每条决策后附一行背景：这个决策是基于什么讨论得出的
- 注意：只有明确的结论性表述才算决策（如"那就定X方案"），
  反复讨论但未拍板的内容不算，请归入讨论过程
- 按主题分组，不要按发言顺序

三、输出「待办清单」，用表格呈现，列包括：
- 事项（具体、可执行的动词开头）
- 负责人（从文本中推断，无法确定写"待认领"）
- 截止时间（文本中提及则写明，未提及写"未指定"）
- 依赖/备注

四、最后输出「未决问题」：讨论过但没有结论、需要下次会议确认的事项。

要求：不要复述寒暄和跑题内容；每条决策和待办必须能在原文中找到依据，
不确定的宁可不写，不要推断编造。

转写文本如下：
{{transcript}}

A few usage tips: if the transcript has clear speaker labels (e.g., "Zhang San:"), the accuracy of owner identification improves significantly; for extra-long meetings, extract in topic-based segments and merge afterward; during review, check the "open questions" first to confirm they truly lack conclusions.

On Cost

If you call via API, this kind of task is extremely cheap — a 60-minute meeting's transcript is usually under 10,000–20,000 characters, and processing it once with a mainstream model costs, per the official pricing page, an order of magnitude less than the value of an hour of human time. That's exactly why I recommend using a sufficiently capable model directly rather than pinching pennies on extraction quality: if the quality of the write-up is unstable, it erodes your trust in the entire workflow.

Final Thoughts

Turning meeting recordings into a decision list and a to-do list is one of the highest-ROI applications of AI in practice: short workflow, non-intrusive, and measurable results. If post-meeting write-ups are dragging you down too, start trying it with your next meeting.

If you need a stable model API endpoint, check this out: https://api.thistoken.ai/register

---

Ready to try it yourself? Sign up at https://api.thistoken.ai/register to get your API key and start building.