First, the Failed Approaches
Turning meeting recordings into a "decision list" and a "to-do list" sounds simple: record → transcribe → feed to AI → receive lists. Many people do exactly this in their first week, then give up in the second. I've seen three typical ways this goes wrong.
Failure #1: Throwing a one-hour recording directly at a general-purpose chat model. The result is either rejection (exceeds the context window) or a fluffy summary like "this meeting discussed issues related to product, marketing, and operations." There's output, but no actionable information. Three days after the meeting, nobody knows what was decided.
Failure #2: Feeding in the raw transcript without cleaning it. Automatic transcription is full of "uh," "um," "you're right, but," off-topic chatter, and greeting small talk. The AI faithfully incorporates all this noise into the summary, and you end up with an 800-word meeting memo mixed with who got stuck in traffic and arrived late that morning. The information density is so low it's basically useless.
Failure #3: Asking for decisions, to-dos, AND risk assessments all in one conversation. The prompt reads "summarize this meeting, as detailed as possible." The model outputs one giant blob of prose—decisions buried in the middle of paragraph three, to-dos scattered everywhere, with no owners or deadlines. This output can't be used directly and requires manual reorganization—at which point the AI's value is reduced to that of a typist.
The common thread in all three pitfalls: treating AI as a stenographer instead of a structured processor. A stenographer records everything; a processor keeps only the information that drives action.
The Right Approach: A Three-Step Pipeline
Step 1: Clean the transcript. After getting the raw text from a transcription tool (Tongyi Tingwu, Feishu Minutes, etc.—check official pricing pages), run a preprocessing pass first: have the AI remove greetings, repetitions, and clearly off-topic segments, keeping only the substantive discussion. The raw transcript of a one-hour meeting is typically 10,000–20,000 characters; after cleaning, it's often down to 3,000–4,000—that's the raw material actually worth analyzing.
Step 2: Structured extraction—run decisions and to-dos in separate passes. The key technique is to ask for only one thing at a time. In the first pass, ask only "what decisions were made in this meeting"; in the second, ask only "what are the action items, who owns each, and what are the deadlines." Running them separately yields noticeably higher accuracy than asking for everything at once, because the model's attention isn't split.
Step 3: Output in fixed formats and land them in your tools. Have the AI output the to-do list as a table (item / owner / deadline / dependencies) that you can paste into Notion or Feishu Base; output the decision list as "one sentence of context + decision + scope of impact," and archive it in your team knowledge base.
A Reusable Prompt Template
Here's the main prompt for use on a cleaned transcript (decision list version). For the to-do version, just swap the task description:
你是一名会议纪要结构化助手。我会给你一份已清洗的会议转写文本。
你的唯一任务:从中抽取本次会议【已明确做出的决策】。
规则:
1. 只收录被与会者明确确认的结论,排除"倾向于""下次再议"等未定项;
2. 未定项单独列入"悬而未决"一栏,不与决策混淆;
3. 每条决策格式:背景一句话 → 决策结论 → 影响范围;
4. 没有决策就如实说"本次会议无明确决策",不要编造;
5. 输出语言:中文。
会议转写文本如下:
【粘贴转写文本】For the to-do version, change the task to "extract action items," change the output format to a table: 事项 | 负责人 | 截止时间 | 依赖条件, and add a rule: "when the text doesn't mention an owner, write '未指定' (unassigned)—do not guess."
Before vs. After AI
Before AI (manual): For a one-hour meeting, you'd spend 40–60 minutes afterward re-listening to the recording or working through the transcript to write up minutes. And most people simply don't do it—meeting conclusions exist only in each attendee's memory. A week later, everyone's recollections start to diverge; two weeks later, some "we agreed on this" item sparks a dispute. For small teams without a dedicated note-taker, this problem is especially fatal.
After AI (pipeline): Transcription, cleaning, and two rounds of structured extraction take about 10 minutes total, with human involvement limited to a final check of owner names and deadlines. For a one-hour meeting, two lists appear in the team group chat within 15 minutes of adjournment: one decision archive, one to-do list that goes straight into the task system. More importantly, the fixed formats make everything directly retrievable—look back three months later, and it's crystal clear which decisions were made in which meeting.
On cost: the token consumption per meeting is minimal (cleaned text is usually just a few thousand characters), practically negligible for indie developers and small teams. Check each model's official pricing page for specific rates.
A Few Practical Tips
First, transcription quality sets the ceiling. For technical meetings with lots of jargon, spend a few minutes calibrating proper nouns first—otherwise "灰度发布" (gray release) might become "会都发布," and no amount of AI cleverness can save that. Second, always keep a human confirmation step on the AI's extraction results—especially the decision list, since a falsely extracted "decided" item can cause more trouble than a missed one. Third, once you've turned this pipeline into a template, you can take it further with API automation: recording file in, two lists out. If you want to call multiple models through a unified interface to run this pipeline and compare extraction quality across models, check out https://api.thistoken.ai/register —register and start building your own meeting-processing pipeline.
The value of a meeting doesn't lie in holding it, but in whether those two post-meeting lists actually land. Let AI handle the recording and structuring, and let people focus on deciding and executing—that's the real point of this workflow.
---
Tired of juggling provider integrations? Register at https://api.thistoken.ai/register and call every model through one base_url.
Хотите попробовать Token.AI?
Создайте API Key уровня проекта, включите каналы в консоли и настройте маршрутизацию, бюджеты и журналы аудита.
注册 ThisToken.AI 并获取 API Key