## The Meeting Is Over, But the Most Exhausting Part Is J...

The Meeting Is Over, But the Most Exhausting Part Is Just Beginning

As someone who has long played the role of "note-taker" on small teams, I absolutely dread post-meeting cleanup.

The standard workflow used to look like this: jot down a few notes during the meeting, then after it ends, replay the recording, listening segment by segment, typing sentence by sentence. For a one-hour meeting, compiling the minutes took an average of 40 to 50 minutes. Worse, the resulting minutes were often a "meeting play-by-play"—who said what, what was discussed, all there, but the truly important things got buried: What did we actually decide? Who is doing what?

The boss reads the minutes and replies, "So what's the conclusion?" At that moment, I realized I wasn't producing information—I was producing a burden.

This isn't just my problem. Among the hidden costs of meetings in small teams, three tasks—transcription, information extraction, and task assignment—often consume more time than the meeting itself. And the longer the meeting and the more participants, the more this cost grows exponentially.

A New Approach: Meeting Ends, Checklist in Hand

Later, I handed the whole thing over to AI and restructured the process into three steps:

Step 1: Transcribe the recording. Use a speech-to-text service (Whisper or the ASR APIs from various cloud providers all work; check their official pricing pages) to convert the recording into timestamped text. An hour of audio is usually transcribed within minutes, and accuracy is already sufficient for Mandarin Chinese scenarios; proper nouns can be corrected via custom vocabulary lists.

Step 2: Structured extraction. This is the core. I feed the transcript to a large language model, but instead of simply saying "summarize this for me," I give explicit structural instructions: extract two types of information from the text—a decision list (conclusions reached by consensus during the meeting, each with its supporting rationale) and a to-do list (specific tasks, each with an owner, deadline, and priority).

Step 3: Five minutes of human verification. Quickly scan the AI-generated checklists, verify task ownership, fill in ambiguous items, then post them directly to the group chat. Within ten minutes of the meeting ending, everyone can see what they need to do.

Throughout this process, AI handles the most time-consuming parts: listening, recording, summarizing, and formatting. Humans only do what AI is worst at: judging right from wrong and providing a safety net.

A Reusable Prompt Template

This is the template I settled on after several rounds of refinement—feel free to copy and use it directly:

你是一名严谨的会议纪要整理助手。请根据以下会议转写文本，
输出两份清单，不要输出其他内容。

【第一份：决策清单】
- 每条格式：决策内容 | 提出人 | 支持依据（转写原文简述）
- 只记录会议上明确达成共识的结论
- 有争议但未定论的事项，单独放在清单末尾"待定事项"里，
  并注明分歧点

【第二份：待办清单】
- 每条格式：任务描述 | 负责人 | 截止时间 | 优先级(高/中/低)
- 负责人或截止时间在会议中未明确提及的，标注"待确认"
- 优先级根据讨论中的语气和紧急程度判断，不确定标"中"

【输出要求】
- 用表格呈现两份清单
- 每条信息必须来自转写文本，不得自行推测添加
- 若转写文本中没有决策或待办，如实说明"本次会议无明确决策/待办"

会议转写文本：
{在此粘贴转写内容}

A few details are worth noting: "Do not infer on your own" is a critical clause. In early versions without it, the model would hallucinate tasks that were never mentioned in the meeting. The "pending confirmation" labels preserve a hook for human follow-up, preventing erroneous information from flowing directly into execution.

Before and After: Let's Do the Math on Time

I kept track of my time before and after the overhaul (based on three one-hour meetings per week):

Processing time per meeting dropped by about 80%. The 2.5 hours saved each week is enough for me to write quite a bit more code. More importantly, there are hidden benefits: the task omission rate dropped significantly. With manual note-taking, content from the latter half of meetings was often skipped out of fatigue; AI doesn't get tired, so a promise made at minute 59 gets treated the same as one made at minute 9. And the structured checklists give real accountability to the age-old problem of "nobody claims tasks after the meeting"—it's in black and white, so nobody can pretend they didn't see it.

As for cost: transcription and LLM API costs work out to a small fraction of the price of a bubble tea per meeting (see each provider's official pricing page for specifics). For a small team, trading a few cents for an hour of labor is a no-brainer.

Three Pitfalls to Watch Out For

1. Don't skip human verification. AI's occasional name-recognition errors and "hallucinated tasks" need a human safety net; five minutes of verification is the safety valve for this pipeline.
2. Transcription quality sets the ceiling. If the meeting room is echoey or people talk over each other, fix the audio capture first, then optimize the prompt.
3. Send the checklists the same day. Meeting memories fade extremely fast. A checklist sent within ten minutes of the meeting ending gets far higher follow-through than one sent the next day.

The essence of a meeting is to reach consensus and assign action—not to produce a play-by-play nobody reads. By letting AI turn the recording into two actionable checklists, people can truly shift from "note-taker" back to "participant."

If you want to build your own AI workflow too, you can start by getting a stable LLM API: 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.