Last Wednesday, a client replied with one line: “Make me a promo video.” I froze again.
At 10 p.m., I was staring at an editing timeline in a Hangzhou café. My client sold handmade bread and wanted a weekend promo, but hadn’t specified the visual style, promotion angle, or publishing platform. I used to ask a round of questions, then keep editing until midnight. Now I handed the brief to a general-purpose AI. It could write a script, but still couldn’t deliver the finished video. I felt a little panicked: the tools keep getting stronger, so why does the work still land on me?
This isn’t a general assistant—it’s a vertical Agent that understands a type of business
Han Sheng, founder of Xiaoying Technology, described a more concrete scenario: a burger shop wants a weekend promotion. The Agent first needs to understand who its customers are and what makes the products worth buying, then select footage, structure the content, and judge whether the finished video can drive in-store visits. That goes beyond “generating a video.” It puts professional judgment inside the workflow.
A vertical Agent is more like a colleague responsible for one defined job. A general assistant is the omniscient butler; a specialist is the operations partner who knows store marketing inside out. For work that is frequent, specialized, and process-heavy, the specialist’s value becomes obvious. The first version usually isn’t delivered as-is either: users still provide feedback, request changes, and only then receive something usable.
How I’d reproduce this approach on a low budget today
Money: I wouldn’t buy expensive software upfront. I’d test the workflow with the AI tools I already have plus a monthly cost of RMB 0–300. Time: I’d build a minimum version in two hours. Technical barrier: coding isn’t required, but I need to explain “how I normally do this” step by step. First step: open my usual AI chat tool, start a conversation named “Short-Video Planning,” and enter: “Break my client’s one-line brief into target audience, core selling points, shot list, storyboard, and revision criteria.”
I made this mistake too: at first, I told AI to “just edit a video,” and all I got was text that looked like a script. What worked was starting with one real client case, then recording every follow-up question and revision criterion until it became a repeatable process.
Here’s how I’d test it at each stage
If I’m just starting out, I’d choose one service I know well, such as shooting event videos for local businesses, and accept only one or two standardized packages instead of rushing to build an “everything for everyone” platform. If I already have one or two clients, I’d turn their briefs into templates: background, goals, shot list, number of revisions, and acceptance criteria. That would make each delivery more consistent. If I’m scaling, I’d build a specialist assistant for the slowest, most experience-dependent part of delivery rather than spreading the budget evenly across every feature.
This tool isn’t for everyone. If my business doesn’t have stable demand yet, waiting is fine; serving a few clients by hand is often more valuable than building a system first. AI coding makes “building it” easier, but the scarce skill is knowing what the client actually wants—and being able to deliver it consistently.