One Month with WorkBuddy: I Need a Secure Ledger, Not 31 Rounds of Chat

Closing up shop at night, I scrolled past a deep dive note on WorkBuddy with a long title: Custom Model Integration, Remote Relay, and Conversation Minutes – 31 Rounds of Practical Essence. Having just finished today's last phone repair order, I had a stack of accessory receipts and customer WeChat screenshots nearby. My first reaction was: if this person had managed a store, they wouldn't just have AI write HTML.

The longer the context, the slower the AI processing speed, and the higher the credit consumption.

I agree with this. I've used WorkBuddy for a month, and the biggest takeaway is that the busier you are, the less you should treat it as a chat partner. In the store, I'm on my feet all day. Repair orders, accessory quotes, refund transfers, supplier statements—it's all piled up in the phone gallery. I want it to organize things, not discuss "why" with me.

Once WorkBuddy can read files, write tables, and sync, it first encounters boundaries. If boundaries aren't set well, the harder it works, the more it resembles an unlocked drawer.

Keep Summaries for Long Conversations, Set Boundaries for Workspaces

There was an interesting approach in that note: using HTML files for conversation minutes, automatically appending summaries per round, previewing anytime, and locating key points in one second. For those who tinker, this is great. I can't write HTML, but I understand what problem it solves. Scrolling through chat history is exhausting. When a WorkBuddy task drags on, earlier discussions about warranty calculations mix with later ones about spreading accessory costs. Asking again two days later means scrolling up. Stores don't have that time.

Now I use it as a ledger entry point. I created a new WorkBuddy workspace called "Repair Shop Ledger." Inside, I only keep two folders: Order Screenshots and Accessory Reconciliation. Other galleries, WeChat files, and customer phone numbers are strictly excluded. I've used the folder structure for two weeks and WorkBuddy for a month; combining these experiences taught me that AI shouldn't be given full access casually. Full access sounds convenient, but trouble arises when things go wrong. Later, I enabled auto-sync only for this workspace, leaving manual sync for supplier contracts. Auto-backup only backs up the ledger, not the entire phone. I've used the Security Center for four weeks; it's annoying normally but saves lives when critical.

Fixed Prompt templates are why WorkBuddy actually started working for me. My template is crude: "You are a repair shop reconciliation assistant. Input is date, platform, order screenshot, or table. Output total orders, refunds, accessory costs, net income, and anomalies. Mark uncertainties. Do not access other folders. Test before automating." I tested it with three days of repair orders, a dozen screenshots, and a CSV exported from a platform. The first version mixed customer-caused refunds with platform subsidies, resulting in a net income discrepancy of over ten yuan. Not huge, but enough to scold it. Later, I added to the template: "Refunds fall into three categories; platform fees only look at net income; list anomalous orders separately." The second version produced a table in about ten minutes with fewer errors. I'm bad at math, so I double-checked with the phone calculator. Books can't be wrong; this step can't be skipped.

Whether table solutions land depends on the entry point. WorkBuddy can turn screenshots, exported files, and WeChat records into a table, making that table the ledger. Yesterday I tried reading fault descriptions from customer WeChat screenshots but got stuck on permissions—which was correct. Turning "screen broken, won't boot" into text, generating an order, deducting inventory, and pushing a quote can't rely solely on model intelligence. Gallery permissions, file permissions, which account can change prices, which can view costs—all must be nailed down first. Otherwise, if it reads wrong, I compensate the customer; if it reads extra files, I lose customer info.

Custom Models Save Money but Shouldn't Be a Barrier for Regular Users

Another part of that note discussed custom model integration, connecting to Alibaba Cloud Bailian, sharing one Key for Tongyi Qianwen and DeepSeek, with zero credit consumption. This claim is attractive to heavy users. I admit that if WorkBuddy's built-in credits burn fast, self-integrating models saves money. But I wouldn't advise novices to do this. I only touched Zhiyuan Interconnect for the first time two days ago and tried DuMate yesterday. My testing shows WorkBuddy's built-in models already handle repair orders, accessory lists, and simple copywriting decently. Custom models involve interface addresses, model IDs, API Keys, maxInputTokens—one mistake leads to a 400 Param Incorrect error. Another Tencent Cloud note configuring MiMo mentioned the difference between id and name, indicating such pitfalls stem from misaligned configuration parameters.

Small store accounting calculates time. Saving credits doesn't save much; saving time is more critical. I can spare at most twenty minutes a day for reconciliation, with no time to research remote relays. Using cloud servers for relay, static IPs, SSH tunnels, and running models on local workstations suits labs, R&D teams, or those truly treating AI as a production tool. My crappy shop relies on an old router for network. If WorkBuddy made me set up tunnels first, I'd have to put reconciliation aside.

But the direction of custom models is useful. It shows WorkBuddy can treat models as swappable parts. For regular users, parts should be hidden behind the scenes. Ideally, options should be "faster," "more accurate," or "cheaper," not forcing me to fill JSON. Someone like me who only knows how to use a calculator doesn't need to know who fits me better between qwen-turbo and deepseek-v4-flash. I just need to know which one distinguishes refunds clearly.

After a month with WorkBuddy, for a store owner like me who can't code, it compresses repetitive office work into templates and turns files into tables. The premise is not going fully automatic immediately, not giving it the entire phone gallery, and not treating "how many models it connects to" as productivity. Those 31 rounds of dialogue, HTML minutes, and remote relays—I'll bookmark them but won't copy them blindly. What I want to copy is the mindset: keep summaries for long conversations, manage context, and give tasks boundaries.

I saw Zhiyuan Interconnect's advanced OA guide mentioning workflow orchestration, data hubs, AI agents, and mobile portal integration. After reading, I still think that's for companies with hundreds of employees. Small stores need an assistant that can read, record, calculate, and lock doors, not a group system. WorkBuddy is still a few steps away from this. Permission docs look too much like engineer manuals, and task minutes haven't directly become ledgers yet. But it's already running.

I don't care about differentiation anymore; I only care if it can smooth out permissions, templates, and minutes. My crappy shop keeps me on my feet all day, and what annoys me most is it being smart without boundaries. If it continues to refine WeChat screenshot reading, order syncing, and permission prompts, I'll keep using it.