跪拜 Guibai
← Back to the summary

Turning a 67-Minute Meeting Recording into Executable Minutes with a Three-Pass AI Workflow

752884fc-fd0e-47e1-85e7-bb605ecbc2c2.png

Contest Topic: TRAE Work Practical Help

Scenario: Organizing cross-departmental weekly meeting recordings, confirming decisions, and tracking action items.

This article: First, it talks about real pain points, then records the practical process, subsequently shows the landing results, and finally extracts reusable experience. All four parts revolve around the same meeting material.

Let's Talk About the Results First

The material to be processed this time is a 1 hour and 7 minute recording of a cross-departmental weekly meeting with 11 participants discussing 4 topics. The original approach was to listen and take notes simultaneously, and when multiple people spoke at the same time, you had to drag the progress bar back and forth repeatedly. Organizing a set of minutes that could be sent to the team usually took 60-90 minutes; after sending it out, people would still ask "who confirmed this decision" and "who is responsible for this next week."

I used TRAE Work to change this into four steps: first, turn the raw material into a structured record, then let it do a single "fact check," and finally, I confirmed sensitive information and responsible persons. What was truly saved was not copy-pasting, but the several rounds of back-and-forth work to find context.

Below is a review of only this one process. I have no product feature introductions, nor did I pile code into the article.

The Original Pain Point: The Meeting Ends, the Work Just Begins

My inputs usually come in three types: the recording file, the agenda from the meeting invitation, and scattered questions collected before the meeting. The three materials are not completely consistent with each other, and the most error-prone areas are three:

I used to cut the recording into several segments, pause when I heard a conclusion, and manually paste it into the document. This method seemed stable, but the most time-consuming part was actually going back to verify: a single conclusion often spanned over ten minutes of discussion, and looking only at the last sentence could easily turn a suggestion into a decision.

Step 1: First, Give TRAE a Checkable Task List

I didn't directly say "help me write meeting minutes." The result of that sentence usually looks a lot like a summary, reads smoothly, but is not suitable for execution. I broke the task down into three parts: input, judgment criteria, and output format.

Raw Requirement Input (Screenshot 01)

image.png

Please read the meeting materials in the current workspace and organize them according to the following rules:

  1. Only record facts that can be confirmed from the original text; do not fill in information that did not appear;
  2. Categorize the content into "Confirmed Decisions, Pending Confirmation Issues, Action Items, Background Information";
  3. Each action item must include the responsible person, deadline, and deliverable; write "To be confirmed" for missing fields, do not guess;
  4. For content with disagreements, retain the viewpoints of both sides and mark who needs to confirm;
  5. Output in Markdown, first provide a one-page summary, then detailed records expanded by topic, and finally an action item table;
  6. When quoting original words, retain the speaker's title and timestamp; if unidentifiable, write "Speaker to be confirmed."

The key to this set of instructions is not "writing beautifully," but setting boundaries the model cannot cross: no guessing, no writing discussions as conclusions, and exposing gaps where fields are missing.

image.png

image.png

Step 2: Let the First Output Expose Problems

The first output is not directly treated as a finished product. I specifically check three things:

  1. Whether the conclusions in the summary can find corresponding evidence in the detailed records;
  2. Whether the responsible persons and dates in the action items actually appeared in the recording or materials;
  3. Whether "To be confirmed" was mistakenly changed to an affirmative sentence.

The most common error in the first output is not typos, but an overly certain tone. For example, if the original words were "we can try it next week," the model easily writes "launch next week." This kind of problem cannot be solved by further polishing; the judgment rules must be written more specifically.

Second Round Correction Instructions (Screenshot 02)

Please perform a fact check on the previous version item by item, do not rewrite the full text:

  • Find sentences in the summary that have no corresponding basis in the original text;
  • Find sentences that write "suggestion/possible/tentative" as "decided";
  • Find action items where the responsible person, date, or deliverable is missing or of unknown origin;
  • Output each problem in the format "Original Sentence | Risk | Suggested Rewrite | Person Who Needs to Confirm";
  • Do not fill in gaps on your own.

image.png

image.png

TRAE Work marks "To be confirmed" and factual risks item by item against the first version

Step 3: Turn the Results into a Document the Team Can Use Directly

After checking, I then ask TRAE Work to output the final version. The final version is fixed into four sections, so team members can jump directly to the part they care about after opening the document.

## One-Page Summary
- What was confirmed in this meeting
- Which issues are still undecided
- Next sync time

## Records by Topic
### Topic Name
- Discussion Background
- Confirmed Decisions (with timestamps)
- Pending Issues and Confirmer

## Action Items
| Responsible Person | Deadline | Deliverable | Status | Basis |
| --- | --- | --- | --- | --- |

## Raw Material Index
- Recording File
- Agenda File
- Related Links

Final Product Instructions (Screenshot 03)

Based on the version that has completed the fact check, generate the final minutes that can be sent to attendees:

  • Retain the four-section structure, do not add generalized content like "Meeting Significance" or "Future Outlook";
  • Every row in the action item table must be traceable back to the original speech timestamp;
  • Use "To be confirmed" uniformly for unconfirmed fields, and list the pending confirmation items separately at the end of the document;
  • Tone is objective, sentences are short, retain necessary context, do not make decisions on behalf of attendees.

image.png

image.png

image.png

image.png Output Final Minutes and Action Item Table

I finally do two manual confirmations: first, replace names, client names, and internal links with desensitized versions publishable for submission; second, randomly check 5 action items and replay the corresponding timestamps. TRAE Work is responsible for speeding things up, I am responsible for signing off.

How to Verify the Results, Not Relying on "Feeling Faster"

Metric Original Process After Using TRAE Work Statistical Caliber
First Draft Time 45 min 8 min Saved 37 min, about 82%
Secondary Check 20 min 5 min Saved 15 min, about 75%
Executable Action Items 2 items 5 items Increased by 3 items
Rework Times 2 times 1 time Reduced by 1 time, about 50%

Note: Not fabricated, it's real .

image.png

image.png

Three Reusable Lessons I Retained

1. First, Stipulate "What You Are Not Allowed to Guess"

When facing dirty inputs like recordings, chat logs, and spreadsheets, the most useful instruction is not "write professionally," but clearly writing the prohibited items: do not supplement facts, do not make decisions for people, mark missing fields. This gives you a handle for the second round of corrections.

2. Separate "Generation" and "Verification" into Two Rounds

Letting the same round write, polish, and check simultaneously makes the model prone to covering up uncertainty for the sake of fluency. The first round only seeks coverage, the second round only seeks risks, and the last round does the formatting. With different responsibilities for the three rounds, the results are more stable.

3. Use "Traceability" to Replace "Looks Complete"

Every conclusion must be traceable back to the original words, timestamp, or file. Adding an extra "Basis" column to the action item table is more valuable than writing another summary paragraph. When someone else takes over, they can quickly review along the basis, and that is reusability.

Common Pitfalls: Why I Didn't Hand All the Work to TRAE

Minimum Template That Can Be Directly Copied

You are my meeting minutes organization assistant. Please organize based on the recording transcription, agenda, and question list I provide.
Only record verifiable facts; do not supplement, do not guess, do not make decisions for anyone.
Output four sections: one-page summary, records by topic, action item table, pending confirmation list.
Action items must have responsible person, deadline, deliverable, original speech timestamp; write "To be confirmed" for missing fields.
First output the initial draft. In the next round, I will ask you to only do a fact check, do not incidentally polish the full text.

Conclusion

The core of this efficiency improvement was not letting TRAE Work attend the meeting for me, but breaking down a repetitive, information-loss-prone task into checkable steps. When encountering similar meetings in the future, I only need to swap the materials, without reinventing the process.

If you also have a task that "you can do but it's time-consuming," it is recommended to start from one scenario, leaving behind the original input, process adjustments, and final product. Only when someone can follow it and do it once is the experience truly delivered.