Zhilv Cloud Scored 9.4, but I Still Dropped That Equity Transfer Agreement into WorkBuddy

High blind-review scores don't mean you can actually use it. Benchmarks test the model; your work depends on the workflow.

Yesterday I came across that 2026 legal AI tool comparison, 8 products, and ZhiLüYun scored 9.4/10 overall in blind review across three document types: labor contracts, equity transfer agreements, and lawyer's letters. The score looks good. After reading it I didn't go try ZhiLüYun—instead I went through that batch of contracts waiting for review again.

From my own experience, this kind of head-to-head benchmark solves "who's smarter," not "what should the output of this step look like." I've used WorkBuddy for two months, and the pitfalls I've hit are basically all on this.

How I set up my workspace

Open WorkBuddy, click Workspace in the left sidebar, then New in the top right. I fix the naming as "client abbreviation-document type-date," e.g. "Hengyuan-equity transfer-1019." Then drag the three documents awaiting review into the input area.

Stop here first, don't rush to hit run.

First thing is to lock the boundaries. Open Workspace Settings, set Allowed modification scope to only the files specified for this task, and uncheck "auto-update referenced files." I've stepped in this pit before. Last time three framework agreement annexes got merged and shifted out of place—it also went ahead and changed the format of an annex I hadn't specified. Since then I changed my habit: save all input files as read-only copies, prefix filenames with _ro_, and cut off its chance to modify them at the source.

Second thing is to write the output contract. This is the core. Don't write "help me review this equity transfer agreement"—write clearly what the output should look like. In Task Description I fix three sections. First section is a risk clause list, with fields locked to clause number, original text excerpt, risk level, revision suggestion; second section is a missing clause check, ticking off item by item against the checklist I uploaded; third section is a conclusion summary, no more than 200 words. Not a single field name can be changed, because later I use batch rename and table merge to reconcile them—if a field drifts by one character, the whole table is ruined.

Third thing is to define the failure path. These two weeks I've been thinking about DAGs—it's just breaking a job into several nodes with a sequence, writing clearly who depends on whom. WorkBuddy's workspace follows the same logic; every node has a failure handling option.

I always choose "stop and mark," never auto-retry or skip and continue. Skip and continue means letting it decide on its own that this step passed, and afterward you have no idea which step was missed. trace_id is the tracking number for each run; when something goes wrong you can locate the specific node, but the premise is that the node has to stop first.

Permissions and division of labor

Team edition workspaces have three permission levels: Owner, Can edit, View only. I generally only give assistants Can edit, clients always View only, and export PDFs with a watermark.

The reason isn't complicated. The intermediate products in the workspace are more sensitive than the final draft—the risk clause list contains my own judgment criteria and concession bottom line, and showing it directly to clients would be a disaster. Also, Can edit permission needs to be thought through together with delete permission. I got burned once—a sentence-by-sentence comparison result got overwritten, but luckily the trace record was still there and I reran it.

Ten minutes before starting work every day

Three fixed things. Take a look at yesterday's run records; open every node whose status isn't "completed"—the kind where you only look at status, not results. Confirm the model version hasn't been auto-upgraded; after an upgrade the same set of prompts produces drift, and wording and field names will change. Copy a fresh template file from the template library for the day; don't edit directly in the previous workspace, or after a while you can't tell which version is the baseline.

My environment is desktop team edition. I hear enterprise edition has finer permission granularity and can isolate workspaces by department, but I haven't tested it, so it may not apply to everyone.

There's one line in that benchmark I agree with: in 2026 legal AI has entered a real-world elimination round. But those eliminated in real-world use are mostly the people who didn't write the output clearly.

I bookmarked that benchmark. The agreement due this afternoon still goes through the old workflow.