My AI Tool List Keeps Growing, but WorkBuddy Is Still What I Open Every Month

I came across that list of the strongest open-source AI coding tools for 2026, divided into four arrays: terminal-native agents, IDE extension plugins, AI-native IDE forks, and complex engineering orchestration platforms. I scanned through it—half I've installed, half I've uninstalled. The list itself is fine, put together quite carefully. After reading it I had just one feeling: the more finely tools get divided, the fuller my bookmarks get, and the only ones I actually open on schedule each month are two or three.

The more complex the orchestration platform, the larger the scope that needs manual checking.

That holds for writing code, and even more so for office chores. Coding agents have tests to catch bugs; a spreadsheet that calculates wrong has no one to catch it—you have to check it line by line yourself.

My monthly reconciliation work goes like this: fourteen files on the desktop, three xlsx exported from three systems, five csv from suppliers (comma-separated plain-text tables, openable directly in Excel), four WeChat screenshots, two PDF statements. Dates come as both text and numbers, and one supplier records amounts in "cents." Before, it took two and a half to three hours. I've used WorkBuddy for two months, and this workflow now sits steadily around half an hour, of which ten minutes is my own review.

First, the directory. Create a 2026-11 reconciliation on the hard drive, with two subdirectories inside, in and out. Dump all raw files into in, don't leave a single one on the desktop. Don't skip this step—the permissions and reruns later all depend on it.

Open the WorkBuddy desktop version—task list on the left of the main interface, New Task in the top right. Click in and it asks what you want to do—don't write a long paragraph in the dialog box; scroll down and find the File Processing template. The template means common workflows are solidified into a form—you just fill in the blanks, no need for it to figure things out itself.

The form has three blanks total: input directory select in, output directory select out, task name whatever you like. The instruction box below is the key. What I wrote was roughly these lines: read all files under in; for xlsx and csv treat the first row as the header; for screenshots and PDFs extract only three items—amount, date, supplier name—nothing else; convert all amounts uniformly to yuan, keep two decimal places, divide by one hundred for those recorded in cents in the original; unify dates into the format like 2026-11-04; leave missing fields blank, don't guess, don't fill them in yourself.

The last line was added only after getting burned. The first run I didn't write it, and one supplier's header was in English—it mapped them to Chinese column names on its own, and even merged two columns while it was at it. The resulting table looked neat, but halfway through reconciling I found the numbers were wrong and spent forty minutes digging back through. The most valuable part of the instruction is spelling out what it's not allowed to do. After adding "leave missing fields blank," the output looks uglier, but every cell can be verified.

Before running, confirm in is clean. This is the step most easily overlooked in daily ops. It scans the whole directory by default, and leftover temp files from the last run get read in too, producing a few inexplicable extra rows. My habit is to empty in after every run and move the raw files under out/archive filed by month—costs ten extra seconds.

On permissions, for single-person use there's not much to say. I gave it just two rules: in read-only, out writable, and no authorization for any other directory on the hard drive. The first time it accesses a directory it pops up an authorization box—read the path carefully before clicking allow, don't casually click "entire disk."

For multi-person sharing I set it up once for a colleague—three people, two read-only and one editable. The two read-only ones can only view results in out and can't touch the instructions; the editable one maintains the template. The reason is simple: the instruction is the only variable in this workflow—if anyone can change it, the output can't be reproduced. Same principle as a code repository.

Going forward, these kinds of tools will probably keep moving toward planning on their own—demos look great. But when it comes to that one table that must be submitted every month, the fancier its planning, the longer I spend checking. I'd rather it be a bit dumber.