After Using WorkBuddy for a Month, I Turned Feishu Meeting Minutes into Raw Material Sources

Last Thursday afternoon we held a quarterly expense review meeting. There were over ten people in the conference room, and the boss said one sentence: "The budget is over; Admin needs to re-categorize last quarter's taxi, dining, and hotel receipts by project." After the meeting, I had 47 minutes of recording, 8 screenshots, 3 contract PDFs, and an attendance sheet on hand. Help! The most annoying part of this kind of meeting isn't the meeting itself, but the mess left behind afterward.

Previously, when organizing meeting minutes, I relied on playing back recordings from a voice recorder and typing manually. Over the past week, I started using a voice recorder with a microphone array for audio capture and just got my hands on Tongyi Tingwu for transcription. It is indeed faster than listening and typing simultaneously. But problems arose: the transcribed output was like a pot of porridge. Who said what—was it about reimbursement, budget, contracts, or submitting materials next Wednesday? Everything was mixed together. I tried throwing it into Notion AI and Claude for summarization; they produced nice paragraphs, but I still had to break down action items myself, verify amounts, edit tables, and send emails.

This week I came across an article on Feishu's official website about their AI Meeting Minutes feature. It mentioned that Smart Minutes can record in real-time, structure key points, automatically break down action items, and consolidate meeting info, to-dos, and quality check results into multidimensional tables, using AI field shortcuts and automation nodes for reminders. I agree with this direction. Meeting data shouldn't be scattered in chat logs, especially for sales, projects, and review meetings. If to-dos are only written in document paragraphs, it basically means no one will claim them.

But I'm doing both admin and cashier duties here. Our company has around twenty-something people. The final landing spot for meeting outcomes isn't a multidimensional table, but a pile of local files, reimbursement forms, contracts, spreadsheets, and emails. If Feishu Minutes stops at "generating a to-do," it's not enough for me. I want it to finish the job for me. This is also the most obvious feeling after using WorkBuddy for nearly a month.

For last week's meeting, I tried a WorkBuddy workflow. I exported audio from the recorder, first used Tongyi Tingwu to convert it into timestamped text, then put the text, invoice screenshots, reimbursement details, and contract PDFs together into WorkBuddy's knowledge base. I've been using the knowledge base for two weeks; initially, I just stored files there, but later realized it's more like building a searchable workspace for AI. Then I gave WorkBuddy a template: extract action items from the meeting, with fixed fields for Item, Owner, Deadline, Amount, Related Project, Required Attachments, and Suggested Email Subject. This template is crucial. I previously posted about WorkBuddy messing up reimbursement summary fields into single lines; later I suspected it wasn't that it couldn't organize, but that the field mapping wasn't pinned down beforehand.

I've switched models a few times in WorkBuddy using models.json and spent some time configuring API Keys, but for meeting organization, the model is just the foundation; the template is the lifeline. At first, I fell into old habits, trying to get it to "summarize the meeting and process all files" in one sentence. The output looked extensive but was barely usable. Later I learned my lesson: let it do one thing at a time—extract first, confirm second, then generate tables and emails. That turned out to be much more stable.

This time, I instructed it to only read the knowledge base without directly modifying files. WorkBuddy ran for about four minutes and produced 11 action items. 8 were accurate, 2 identified owners as departments instead of individuals, and 1 broke "submit contract attachments by next Wednesday" into "contract approval." In the amount fields, taxi and dining receipts mostly matched, but one hotel receipt mixed tax-inclusive and tax-exclusive amounts. My testing shows that relying purely on AI for the final version isn't realistic yet, but compressing roughly two hours of organization into thirty minutes of review is already a blessing for office workers.

What really made me feel the difference between WorkBuddy and other meeting tools was the subsequent steps. I fed the confirmed action items back into WorkBuddy, asking it to draft emails by owner, rename attachments as "Date-Project-Type-Amount," organize reimbursement-related screenshots into a temporary folder, and simultaneously generate a CSV for Meike (expense management software) import reference. I've used Meike for two weeks; the fields and WorkBuddy's export table still need manual adjustment, but at least I don't have to copy-paste while listening to recordings. I also tried PPT generation, asking it to compress meeting highlights into three slides: one for overrun reasons, one for action items, and one for pending confirmations. The result was mediocre—still too much text—but the framework was usable.

However, the stronger tools like WorkBuddy become, the tighter permissions must be. I only seriously tackled permission modes in the last week. Before that, reading enterprise deployment articles made me nervous because I was constantly uploading invoices, contracts, and reimbursement forms—it was basically running naked. Now my approach is simple: financial raw files are read-only; generating drafts is okay, but direct moving or overwriting is not; bank statements, supplier contracts, and employee attendance raw sheets are blacklisted; all external actions, including sending emails, editing shared tables, and batch renaming, go through approval mode. Amazing. I used to think automation meant clicking fewer mice; now I realize automation fears skipping brainpower the most.

So if forced to compare, Feishu Smart Meeting Minutes is better suited for teams where meetings stay within the Feishu ecosystem, collaboration is online, and to-dos need to be consolidated into multidimensional tables for continuous tracking. Its strength lies in connecting the live meeting, records, reminders, and data tables, which feels very smooth for project teams and meeting-intensive organizations. WorkBuddy is better suited for someone like me who faces local files, reimbursements, contracts, emails, spreadsheets, and archiving after the meeting ends. It doesn't just try to make the meeting "clear," but tries to get the meeting "done." If an AI office tool can only generate minutes but cannot turn those minutes into files, tables, emails, and approval actions, it will likely end up adding another layer of review work for employees.

Of course, WorkBuddy isn't omnipotent. Over these two days, what scares me most is its overconfidence in understanding context. If someone says in a meeting, "Handle that reimbursement however you see fit," it might actually generate a handling plan. But behind "however you see fit" lie financial rules, budget categories, and the boss's temper—AI doesn't know that. Field templates can save part of it, external rule bases can save part, and approval modes provide a safety net, but you can't expect it to take responsibility for errors directly. Especially in finance, it's better to be slow than to let it act on its own initiative.

Now I treat Feishu Meeting Minutes as the source material and WorkBuddy as the executor. Recordings, transcriptions, minutes, and multidimensional tables can all exist, but what I ultimately look at is: who does it, where are the files, are there email drafts, do reimbursement fields match, and is archiving correct. Meeting records solve "knowing what was said," while WorkBuddy solves "knowing where to act next."

This post might seem picky about tools. But I genuinely believe that the next battle for AI office tools isn't about who can write notes that look more like notes, but about who can catch the miscellaneous tasks after the meeting. When you organize meeting minutes, do you care more about transcription accuracy and pretty bullet points, or whether it can directly generate to-dos, tables, and emails?