One Month with WorkBuddy: Not as Codex, But as a Construction Site Clerk

It was past 8 PM, the AC in the project office wasn't working well again, and I was scrolling through my computer when I came across an article comparing the approaches of Codex and WorkBuddy. My first reaction was to push aside the pile of stuff next to me: rebar delivery plans, safety inspection records, several Excel quantity calculation sheets, and material approval forms requested by the supervisor. Another day covered in dust. The article said Codex uses multi-Agent parallelism and is suitable for large-scale refactoring, which sounds impressive. But I'm someone who handles documentation; my job is to turn a pile of dusty paperwork into something deliverable.

I've been using WorkBuddy for about a month. In the first few days, my mindset was very simple: skip the fluff, can it help me calculate quantities? Later, I actually used it for table cleaning, weekly report organization, and file archiving, and I hit some pitfalls too. At the end of August, I wrote about it misreading garbled text in drawing parameters; last week, I wrote about it losing formatting when converting weekly reports and progress tables to PPT. Looking back now, I think WorkBuddy is more like a construction site document clerk. If you give it rules, it works. If you throw a bunch of headless, tail-less info at it, it gets confused too.

The article compares Codex and WorkBuddy separately: one follows the developer route, the other the general workplace user route. I basically agree with this judgment. Codex's ability to modify files in parallel, run tests, and view diffs definitely has a market in software engineering. On the construction site, what annoys me most is inconsistent standards in material planning tables. For the same rebar specification, the crew reports A, the budget reports B, and the site photos show C. There's a lot of dirty data in Excel, and quantities are hidden in PDF notes.

What I found truly useful about WorkBuddy is its ability to connect fragmented tasks in office workflows. Based on my testing, handling Excel quantity tables has become smoother than at the start. Previously, manual merging took half an afternoon; now I can get a first draft in about ten minutes. It organizes according to the column names I provide and can pick out duplicates, null values, and abnormal units. To put it bluntly, this thing is faster than my crappy computer—at least I don't have to click line by line. It still fails with parameters in drawings, often mistaking notes for parameters. I don't expect it to draw CAD diagrams for me; SolidWorks and CAD output tools are for that. It's more like helping me organize documents clearly.

I eventually realized that WorkBuddy's design philosophy isn't the same as Codex's. Codex is like throwing multiple Agents into isolated environments to work in parallel and merge results at the end. WorkBuddy is more like a task chain, proceeding step by step, one action after another. It has three modes: Ask, Plan, and Craft. I mostly use Plan. First, let it generate a plan, I check if the steps are correct, then let it execute. This is more stable.

It also has so-called SOUL.md, USER.md, and MEMORY.md layers of memory. I thought this was mystical at first, but using it in projects saved effort. I write common standards into USER.md, such as concrete grades being C30, C35, C40, no spaces in rebar specs, and dates unified as YYYY-MM-DD. MEMORY.md records project names, sections, and common headers. This way, next time it processes similar documents, it doesn't have to guess anew every time. It doesn't truly understand; it just lowers context costs.

Another key point of WorkBuddy is embedding. It can work directly within WeCom, Tencent Docs, QQ Mail, and WeChat Mini Programs. Construction site documents are often scattered across WeChat groups, email attachments, and local folders. I had it categorize Excel and PDF files from several folders, batch rename them, and create directories by date and building number. The result was faster than me clicking the mouse myself. But this also brings a limitation. It mainly works within the apps it integrates with and local files; it can't freely control the screen like some desktop Agents. This might not be free enough for developers, but for document organization, it reduces risk.

After using it for a month, I learned to let it do only one thing at a time. Don't come in saying "handle all the project documents," or it will make a mess, and the format will still be wrong. Templates for headers, column names, units, and date formats must be provided upfront—the more specific, the better. Don't force complex format conversions like PPT/PDF to tables; on my end, it's best to pre-convert or import single files. Results for quantities, specifications, and dates must be manually reviewed; one mistake means rework. It can handle organizing, archiving, cleaning, and summarizing, but don't expect it to replace professional design and precise quantity calculations.

WorkBuddy's strength is breaking office tasks into executable steps. This differs from Codex's parallel engineering route. Codex is suitable for large project refactoring; WorkBuddy is suitable for daily repetitive document work.

I've said before that doubling image generation speed is just a surface change; enterprise-level implementation cares more about details like format conversion, data cleaning, and file archiving. I still hold this view. Many people on forums love seeing "generate PPT with one sentence" or "one-click renderings," which sounds exciting. But in the project office, if PPT loses formatting, table columns shift, or PDF notes are missed, humans still have to clean up the mess. In last week's post about converting to PPT, many comments mentioned losing templates, fonts, and charts. I've encountered this too. Eventually, I learned my lesson: let it output structure and text first, then fine-tune formatting in Tencent Docs.

I guess the areas where WorkBuddy needs to improve next are these. Drawing structuring needs to be more accurate; it shouldn't mix notes, descriptions, and legends into parameters. Table template fidelity needs to be better, especially for Excel merged cells, filters, and formulas—they break easily. Error messages need to be clearer; don't just give results, tell me which row and column are uncertain. Permissions and versioning also need to catch up; project documents involve crews, subcontractors, and supervisors—who changed what should be traceable.

I won't blindly praise it either. Right now, it's like a diligent document clerk, far from being a Chief Engineer. If you throw drawings, tables, emails, and documents at it, it can organize 70-80% usable output. The remaining 20-30% still needs human review. Quantities on construction sites cannot be wrong. If the material delivery plan is off by two trucks, the site goes chaotic. WorkBuddy saves me time, that's true. But I dare not let it make decisions for me.

The question at the end of the article is interesting: Will AI Agents move toward becoming an all-capable development foundation, or toward office products usable by everyone? I care more about the latter. Most project offices don't need an Agent that can modify code; they need an Agent that can clearly organize tomorrow's briefing materials, progress ledgers, safety records, and material approval forms. I don't think WorkBuddy is going down the wrong path. In the future, if it can automatically grab records from WeCom, update ledgers, and generate approval forms, permissions should be granted gradually, and review must remain with humans.