Building a Personal Workbench with WorkBuddy: Map Out Failure Paths Before Writing Prompts

The place where an AI workbench most easily crashes is that when it makes a mistake, you have no idea which step went wrong. Whether the model is smart is secondary.

Today I came across an article about landing an AI personal workbench, which mentioned DAG definitions, node execution, failure retry strategies, and specifically talked about trace_id tracking, saying any layer's problem can be located to the specific step. The parameters section left an impression. Running a 14B model on the intranet, input-output cap 4096, temperature 0.2, the author says this config suits deterministic tasks like document summarization and information extraction. Sounds like it's written for engineers, but the thinking applies to WorkBuddy just the same.

My setup is like this. WorkBuddy used for about two months, workspace used for a while, three-plus weeks. What follows may not apply to everyone, especially the team collaboration part.

I've seen too many people, including me last month, treat the AI workbench as a wishing well. Throw in "help me review these three contracts" and wait for results. The result is misalignment. Last Wednesday I was handling three supplier framework agreements, and after merging the attachments, similar cases, and review opinions, in the output table Company A's clauses got attached under Company B's name. I checked for a long time, and the problem was I hadn't defined which step should produce what.

There's a line in that article I agree with. Long-running workflows should be split into multiple transaction segments, updating progress after each node completes, so the user can clearly see which step it's stuck on, rather than facing a button that spins forever. This experience detail directly determines whether the tool can be accepted in real business.

Whether the failure path is clearly written directly determines success or failure.

After failure, retry per strategy: failures from network jitter retry 2 times, business validation failures don't retry and throw an error directly.

Put into a legal document workflow, this translates to: format errors can be rerun, factual errors must stop and let a human look. My judgment is that a tool like WorkBuddy performs unstably in multi-stage workflows, stuck on process decomposition and boundary definition, not much to do with model capability. Machine-generated lists can only be drafts, the material correspondence must be manually verified. This point my thinking hasn't changed; what's changed is I now proactively write validation points into the prompt, no longer waiting for errors to go back and find them.

Specifically, this is how I set it up. Open WorkBuddy, click Workspace on the left, create a new one, name it something like "Supplier Contract Review-2026Q4" with date and purpose, don't call it "New Workspace 1" — in two weeks you yourself won't remember what it's for.

After entering, click Add Material, drag the PDF in. Note, only put one type of material at a time. Three contracts are three contracts; similar-case search results go in a separate workspace. This is the first pitfall — mixing them guarantees misalignment.

Then the key step, configure output. I don't use its default summary; instead I lock the fields in the prompt, six items total: contract name, clause number, original text excerpt, summary (no more than 80 characters), risk level (high/medium/low), source basis. The benefit of locking fields is the output can go directly into my evidence chain ledger, no second-round organizing. Parameters like temperature I can't see; what I can control is the constraints in the prompt, so the constraints must be written more rigidly than you'd think. I've suffered on this — write it loose and it improvises.

On permissions, the workspace has Share in the top right. My habit: assistant set to Can edit, only let him add materials and annotate; client set to View only, send him the export link. Who changed what, there's a record in the workspace, way better than my old back-and-forth by email. If permission boundaries aren't locked, both the model and collaborators will act on their own — this is an old problem.

Daily maintenance is just two things. Clean intermediate products in the workspace once a week, especially the previous round's output — keeping it will pollute the next round. Every time you change the prompt, run a sample first, confirm the fields haven't drifted before going to batch.

These two days I've been trying MCP, wanting to connect it into WorkBuddy so similar-case search results automatically land in the workspace. Just started, hard to draw conclusions yet, I'll write once it runs smoothly.

Next time before using WorkBuddy, spend ten minutes writing down what this step should produce and what to do if it goes wrong, then start. The rework time saved is enough for two cups of coffee.