I rebuilt Bee Home-School's grading loop with WorkBuddy
Community Discussion · General

I rebuilt Bee Home-School's grading loop with WorkBuddy

Teacher LinTeacher LinOct 62026/10/06 471 views

I've always felt that tools that truly save time are mostly doing bookkeeping for you.

A couple of days ago I came across the Bee Home-School AI intelligent homework grading management system; its feature list is very complete. Textbook-synced dictation and reading-aloud, essay marginal comments plus full-text polishing, automatic collection into a class mistake notebook, learning analysis marking accuracy by knowledge point, and a smart reading all-in-one machine that scans, prints, and carries grading marks on the original paper. It's said to finish concurrent grading for a whole school in a short time, with students needing no terminal and teachers needing no manual operation, and reports coming out by themselves.

As I read, I laughed. The logic of this system is beautiful, but what it actually solves is two things: the grading itself, and the accounting after grading. The first thing I can't do; our class has over fifty essay notebooks, and I don't trust a machine to recognize handwriting. The second thing, I used to not do at all.

So this time I didn't compare who grades more accurately. I used WorkBuddy to build the post-grading accounting along the same lines. My environment is the Windows version of WorkBuddy, using data from our class's third essay this semester; it may not apply to everyone.

This kind of work is most easily messed up by throwing the raw material directly at AI and letting it summarize. I stepped in this pit before. Last semester I tried photographing the whole class's essays, converting them to text, and dumping it all in to have it analyze our class's writing problems. What came out was beautiful enough to stick on the blackboard, but not a single sentence could be pinned on a specific student. Later I changed my approach: do only one thing at a time.

Specifically, it went like this. First create a workspace in WorkBuddy; I named it "Class 5(2) Essay Review." This workspace is like a fixed folder, and all later tasks hang under it. Then go into the task panel; on the left there's a column for "Upload Files." Put this round's raw materials in: one Excel of the student roster, and one text file of essays I made by image-to-text, with each essay manually prefixed with a line containing name plus title. That line is crucial; everything later depends on it for alignment.

Next, create the first task. Click "New Task" in the upper right and choose "Table Processing." I wrote the prompt very rigidly: read the essay text, split into one row per name, output four columns: name, title, word count, one-sentence summary of content, no evaluation. It ran for about ten-plus minutes and produced forty-some rows. I don't let it evaluate at this step because once it starts evaluating, all later classification will skew.

The second task is the key one. I created a new table and moved over the mistake-notebook logic, except ours is an essay version. I had WorkBuddy tag each entry by the dimensions I gave, just three: whether the opening is boilerplate, whether there are concrete examples, and whether the ending goes off topic. In the prompt I wrote the criteria for each dimension clearly, for example boilerplate means the opening contains sentences like "in my life" or "everyone has." This step ran for about twenty-plus minutes.

After it finished I got a table where each row is a student and each column is a problem. This is what Bee Home-School calls the mistake notebook and weak points at a glance, except my version uses dimensions I defined myself, rough but handy. I exported it to Excel and archived it in the workspace.

It wasn't over yet. The third task is what made me feel this process stood up: grouping by problem. I had it pick out the students marked yes in the boilerplate-opening column, generate a separate roster, and attach each one's original first sentence. On Friday's review class I taught from this roster: whose opening needs changing and why, all visible on one screen. This is much cruder than Bee Home-School's one-screen display of answer situations, but it's a dimension I picked myself, so I feel grounded when teaching.

Two pitfalls to mention. The first is file naming. At first I put both the essay text and the roster in the workspace root, and as a result it occasionally read across files and stitched two students' content together. Later I created two subfolders in the workspace, one called Raw Materials and one called Generated Results, and wrote the paths explicitly in the task, and it never happened again. The second is permissions. I set this workspace to visible only to myself, because it contains students' names and original text, and I don't want it synced anywhere. If someday I need to work with teachers in the same grade, I'll create a shared workspace and put only the de-identified result tables there; raw materials won't enter the shared area. I figured out this split gradually too; at first I put everything together to save trouble, then thought about it and decided it wasn't appropriate.

Daily maintenance is just one thing: after each review, export the whole workspace for that round and save it locally, then archive it in WorkBuddy, don't delete it. By the end of the term, these tables stacked together form a fairly solid class writing archive, more reliable than recalling from my own impressions.

As for Bee Home-School's all-in-one machine, cloud question bank, and district-level horizontal comparisons, I probably won't use them, and there's no need to. But it reminded me of one thing: the value of grading is half in the grading, and half in whether the accounting after grading can be made clear. The part I couldn't account for clearly used to just get glossed over. Now at least our class's essay accounting can be laid out item by item.

I guess in the next year or two, the competition among these school systems will shift from whether grading is accurate to whether the accounting is clear, because grading itself will sooner or later be commoditized, while the ledger every teacher wants looks different.

2 replies

?
Ctrl + Enter to reply
Crypto Dropout

Your "porting the wrong-answer-notebook logic into an essay version" is way too real, but all three dimensions are self-defined — swap in a different person and you have to relabel everything. That's the hardest part to tokenize.

Mo Mo
Mo MoOct 8

If those three tagging dimensions of yours require manually adding a name line to every essay, going through forty-some books would be more tiring than the grading itself.