After a month with WorkBuddy, I integrated file organization into expense reports using a 'return to place' approach

Today I came across a post: xntj-ai released a local file auto-organization service called guiwei. It's MIT open source, has only 16 Stars, but the README is pretty intense: drop files into an inbox folder, let the LLM read the content, rename, categorize, and build a reversible index. Documents/images use OpenAI-compatible APIs, while audio/video uses Gemini.

Help me out here—this description basically recounts every pitfall I fell into with WorkBuddy last week.

After one failed desktop cleanup, I realized WorkBuddy isn't bad at organizing; I just didn't give it a rollback plan

I handle both admin and cashier duties, so my desktop is always piled high with invoice screenshots, contract PDFs, attendance sheets, and meeting minutes. On September 1st, I wrote a post asking if my rules were misconfigured when WorkBuddy automatically organized my desktop files. At the time, I asked it to categorize by "Reimbursement," "Contracts," "Spreadsheets," and "Others." The result? It mixed taxi receipts and hotel folios into "Spreadsheets" and dragged a contract attachment screenshot into "Reimbursement." My desktop ended up messier than before.

Last Thursday, after our quarterly expense review meeting, I had over forty minutes of recording, a few folios, one contract PDF, and an attendance sheet. The most annoying part of this scenario isn't the meeting itself, but the mess left behind afterward. Previously, I relied on replaying voice recorders and typing manually. Over the past week, I started using voice recorders and microphone arrays for capture, transcribing via Tongyi Tingwu (Aliyun), and then letting WorkBuddy process the transcripts and attachments.

After that incident, I was actually hesitant to touch WorkBuddy again. I previously felt it was only suitable as a draft generator, not without reason. It's too good at "looking plausible," but lacks stable understanding of financial fields, document entities, and action boundaries. When you ask it to organize files, it assumes you're doing personal tidying; but I'm doing reimbursement archiving, where each file is tied to projects, budgets, approvals, and payment statuses.

Seeing guiwei made me realize the problem wasn't that WorkBuddy couldn't organize at all, but that I hadn't given it three things: an inbox folder, an index, and reversible actions.

Folders should just be dumb storage; retrieval value lies in the index.

I really resonate with this sentence. What admins fear most isn't having too many files, but not knowing which version is the real one when searching. So I changed my WorkBuddy workflow to look like this:

• First, fix an "To Be Filed" inbox folder; don't touch the desktop directly.
• WorkBuddy only reads, doesn't move. It outputs a list first, with fields following my template: Date, Type, Entity, Project, Amount, Suggested Folder, Needs Manual Confirmation?
• For invoices, folios, and contract PDFs, I have it generate suggested filenames using the format Date_Type_Entity_Subject, e.g., 20260903_Invoice_TravelProject_HSR_Ticket.pdf.
• For recordings and transcripts, I first run the text through Tongyi Tingwu, then feed the text and attachments together to WorkBuddy to generate meeting minutes, action items, and pending reimbursement fields.
• Finally, I confirm the list, then let it execute the moves, keeping an organization log.

In my testing, processing about 20+ files, roughly half went straight into the correct folders, 40% needed field corrections, and the remaining two or three were skewed photos of invoices or missed remarks. This is much more reliable than just telling it to "auto-organize" from the start. It didn't get smarter; I just stopped letting it act as the judge.

There are also shortcomings with WorkBuddy here. It doesn't have built-in financial knowledge. Last time, the summary fields for reimbursements kept merging into a single line. I initially thought it was a prompt issue, but later realized my field mapping was too scattered. WorkBuddy doesn't understand that "Supplier Name" and "Payee Entity" aren't the same thing in certain contexts, nor does it know that "Project Name" and "Expense Attribution" might need to be split into two fields. So now, I create a template table first, copying fields from Chengbao Reimbursement and Meike systems, and then have it extract data according to the template. Only then is the output quality stable.

But what truly makes me willing to keep using WorkBuddy isn't file renaming. It's its ability to connect file organization to actions. After organizing a batch of travel invoices, I have it output a "Pending Materials" table, listing missing folios, approvals, or project codes. It can also break down Tongyi Tingwu meeting transcripts into tasks and generate a weekly report draft for my boss. Is this a godsend for office workers? Not exactly, but it saved me half a day of verification. Previously, AI tools only gave me text, making verification costly; now, at least it turns text into verifiable tables and to-do lists.

Compared to guiwei, WorkBuddy's strength isn't "filing," it's "execution"

guiwei's approach is very clean: local service, reversible, indexed, directory.md. It suits individuals and small teams, as well as those willing to tinker with scripts. Using OpenAI-compatible APIs for docs/images and Gemini for audio/video is a pragmatic division of labor. However, it's essentially still a file steward. You throw files in, and it helps file them, rename them, and build indexes. If you ask it to generate reimbursement forms, break down action items, or write weekly reports, it won't necessarily be smooth.

WorkBuddy is the opposite. It's not a local service, deployment isn't as lightweight, but its office actions are broader. I use it for documents, spreadsheets, PPTs, data analysis, and email drafts, and it handles them all. What really changed my mind was starting to treat it as an "action layer" rather than an "organization layer."

If looking strictly at file organization, guiwei's design is more mature. WorkBuddy should learn from the reversible index feature. My current organization logs are something I require myself; they aren't provided by default by the product. Without rollback, AI auto-moving files is a dangerous operation. Especially in administrative scenarios, if a contract attachment is moved incorrectly, auditors will later ask, "Why isn't the version here?" I dislike products that sell "fully automatic AI" as a feature. Full automation isn't the goal; controllability is.

If looking strictly at office execution, WorkBuddy is stronger. It can extract fields from files, generate tables, write emails, break down tasks, and output PPT outlines. Its problem lies here too: capabilities are too scattered. If you give it too many tasks, it gives you a result that looks complete but has drifting internal logic. So I have a principle when using WorkBuddy: Let it do only one bounded task at a time. For example, today only organize travel invoices, tomorrow only map reimbursement fields, and the day after, write meeting minutes. Don't let it read contracts, rename files, and summarize meetings simultaneously.

I also learned a lesson regarding permission modes. Initially, I connected some knowledge bases via API Key, thinking it could stably reference company templates. Later, I found it would confuse old reimbursement policies with new ones. Now, I tag templates in the knowledge base with status: Effective, Deprecated, Pending Confirmation. When WorkBuddy retrieves fields marked "Pending Confirmation," it puts them in an exception list instead of executing directly. This method is crude, but it works.

An actionable step for others using WorkBuddy

If you're also using WorkBuddy to organize files, I suggest not rushing into full automation. Start with a dry run on 20 files. Have it only output a list without moving any files. The list must include at least four items: suggested filename, target folder, extracted fields, and reason for exceptions. Confirm manually once, then solidify the rules into a template. The template doesn't need to be complex; for reimbursement scenarios, having "Date, Amount, Supplier, Project, Expense Type, Attachment Status" is enough.

After using WorkBuddy for a month, my judgment of it has shifted slightly. Previously, I thought it was just a draft generator; now I think it can serve as a semi-automated processing line, provided you give it boundaries, templates, and rollback mechanisms. guiwei showed me not a stronger competitor, but a clearer design principle: Files themselves don't need AI performance art for organization; indices and actions are what deserve to be preserved.

Only AI that lands on actions deserves to stay in your workflow. But before landing on actions, you must make it tell you how it plans to land, and how to retreat if it's wrong.