After a Month with WorkBuddy: Splitting Scheduling into Three Stages

After night shifts, the thing I fear most is a phone notification saying, "Send the schedule." It's not that I can't send it, but every time I have to reopen the hospital system, export CSVs, merge them in Excel, and figure out how to send it out. After using WorkBuddy for a full month, my conclusion is straightforward: it can get the work done, but don't treat it as a wish-granting machine. Optimize first, then talk about efficiency.

Identify bottlenecks first, then let WorkBuddy do the heavy lifting

Today I came across content on Arthas performance tuning, discussing Java diagnostics, tracing method durations, and using flame graphs to find flat-top bottlenecks. I don't write code, but I think office tools can borrow this mindset. Previously, I treated WorkBuddy as an all-purpose assistant, telling it to "organize the schedule, generate Markdown, and email it to the team leader." The result was messy table formats and emails that never arrived. Later, breaking it down, I realized the problem wasn't in one spot.

Diagnostic tools like Arthas put it simply: look at duration first, then find the bottleneck. Don't guess.

My testing showed the slowest part wasn't AI understanding, but unclean input. I've used the hospital system for four weeks; I see the raw exported CSVs daily: inconsistent column names, empty rows, shift notes mixed into the last column, sometimes even a total row. Throwing this directly at WorkBuddy takes about five to six minutes to process, but the output has misaligned rows, requiring me to spend thirty seconds to a minute fixing it, which is actually more annoying. Later, I started doing three small things in Excel first: fix headers to Name, Date, Shift, Notes; delete empty rows; split "Night Shift" and "Day Shift" from the notes into separate columns. Then giving the cleaned CSV to WorkBuddy, it categorizes night shifts, day shifts, and rest days, generating a Markdown table in about two minutes. Misalignments basically disappeared.

WorkBuddy's bottleneck is often not in the AI, but in the input format. Initially, I always wanted it to infer rules from dirty data with one sentence, but later realized that was wrong. How good a tool is depends on one thing; mastering basic usage before letting it do heavy lifting is another. In the workflow automation features, I now only let it do three fixed steps: read, categorize, output. The email step runs separately. I've only been using email integration for a few days, so I dare not claim stability, but manually exporting attachments is definitely more controllable than scheduled tasks.

I also found a very simple optimization: file names. Previously, I had it output "Schedule Result," repeating daily. Later, I changed it to "Schedule_Week_Version," making it less likely for WorkBuddy to mix up historical tasks when reading them. I don't have a dedicated performance panel, but fields like status, time, and result path in the task list are enough to judge whether it didn't run, ran incorrectly, or finished without delivery. This is the same idea as Arthas looking at duration before bottlenecks—don't blame the model for being dumb right off the bat.

Schedule drafts are 80% ready, but human relations can't be outsourced

I mentioned before that industry expert features can get 80% of the work done. After using WorkBuddy for a month, my judgment remains the same. But the remaining 20% after that 80% is the most grueling.

Our department's scheduling looks like a spreadsheet, but it's really about relationships. Who just returned from maternity leave, who needs to pick up kids on Wednesdays, who cannot work the same shift as whom, who must take compensatory leave after consecutive night shifts. WorkBuddy can run by rules, but it won't bear the consequences for me. Now I have it output three fixed tables:

• Shift distribution table: Night, Day, Rest, summarized by person.
• Consecutive shift risk table: People with more than two consecutive night shifts highlighted separately.
• Handover Markdown table: Easy to paste into tablet handover documents.

My testing shows that manually creating next week's schedule used to take about forty minutes. Now, WorkBuddy's draft plus my manual final review takes about twelve minutes. The saved time isn't for resting; it's for verification. AI scheduling must have manually set rules and retain final review authority. I've used both the hospital system and a certain AI medical record system for four weeks; each has its boundaries. I've only used AI diagnosis for a week, so I absolutely won't let it make judgments for me. WorkBuddy is the same; it handles back-office tasks, not core clinical records.

A few days ago, I wrote that WorkBuddy's remote configuration didn't work, unsure if the computer wasn't awake or if I hadn't bound it correctly. After continuing to try, I ultimately didn't make remote access the primary path. After a night shift, I don't want to gamble on whether my phone can wake up my computer. Now I run it locally before leaving work, exporting results directly to CSV and Markdown, and sending attachments manually. Scheduled tasks show ACTIVE, but emails aren't received. I'm testing this over the next few days; maybe path, permissions, or sleep settings are mismatched. Whether a tool can work remotely is one thing; whether I dare to let it send schedules remotely is another.

Here's another pitfall: WorkBuddy messed up the table formatting after processing. I initially thought it didn't recognize Chinese column names, but later found hidden spaces and line breaks in the original table. After cleaning, its Markdown output became very stable. This pitfall is trivial, but people working night shifts don't have the energy to deal with "formatting is broken again." So I hardcoded the rules: No empty cells in the name column; Shift column only allows Night, Day, Rest; Notes column excluded from statistics. Better to spend two extra minutes on preprocessing than to let it guess.

WorkBuddy suits standardized people, not lazy ones

This post isn't praise nor a warning against use. After a month, I think WorkBuddy suits two types of people: those with fixed processes, and those willing to standardize inputs. It also suits those who want to outsource repetitive document, table, and file organization. It doesn't suit those expecting "one-click takeover," especially not in healthcare scenarios requiring signatures, human touch, and clear accountability.

Looking cross-industry, AI office tools aiming straight for the top tier often crash and burn. If WorkBuddy only wants to be an "all-purpose assistant" but hasn't sorted out basic paths, email, permissions, and formats, users will encounter the awkwardness I faced days ago: Task executed, result undelivered. It's not unusable; it's just that it shouldn't be used recklessly.

I've set hard rules for WorkBuddy now. Input must be CSV with fixed column names; Output must include date and filename; Files involving personnel info don't go to public cloud drives; All schedule drafts require manual final review; Email integration stays manual until stable. These rules seem like shackles, but they're actually integrating it into my workflow.

Another practical aspect: it shortens the transfer between Excel and Markdown. I've used Excel for three weeks; previously just for making tables. Now Excel handles cleaning, WorkBuddy generates handover docs, and tablets handle viewing. The chain is shorter, but every segment needs someone watching. The more "automatic" a tool seems, the clearer humans must draw the lines of responsibility.

Don't call me a white angel, and don't expect a tool to clean up all the mess after a night shift. WorkBuddy currently helps compress my schedule draft to around ten minutes; the rest is still edited by me. It saves trouble, but it can't sign for me.