Editor's note: This post is adapted from a talk I prepared for the unveiling ceremony of the OPC Ecosystem Alliance's innovation center at the JD (Zhengzhou) Tower. The original draft used "AI team" as a speaking metaphor; this post further distinguishes between workflows I've actually tried, projects still being iterated on, and scenarios used only to illustrate a method.
My name is Zhang Yukui. I studied clinical medicine — a non-technical background.
I can't program independently, and I don't have a technical team. But I've been trying to do one thing: with AI's help, can I take content, information, and operations work that would otherwise require repeated manual effort and break it down into a set of AI-assisted workflows that can collaborate with each other?
I call these workflows my "AI team."
That phrase doesn't mean AI can actually replace a full role, and it doesn't mean one person can conjure up a mature company out of nowhere. What it describes is a way of working: the human sets direction, standards, and judgment, while AI takes on part of the repetitive execution.
How This Differs from Just "Using a Tool"
A lot of people use AI to have it edit an email, summarize an article, or look up some information — a one-off task. Once the task ends, the conversation ends too.
Team-style collaboration puts more emphasis on stable division of labor:
- What information enters the workflow;
- Which step AI is responsible for;
- What standard is used to check the results;
- When a human must step in;
- How errors get caught and logged.
Rather than chasing an all-powerful AI, it's better to set clear boundaries for each type of task.
First Team Member: The Content Assistant
While running a WeChat Official Account, I realized that the parts I truly need to own myself are the viewpoints, the case examples, and the judgment calls — while gathering material, drafting structure, writing first drafts, and formatting involve a lot of repetitive work.
So I first started treating AI as a content assistant:
- Feed it my past articles so it understands my usual tone and structure;
- Be clear about who this piece is for and what it's trying to say;
- Have AI generate a first draft;
- Check the facts, cut the boilerplate phrases, and add in my own firsthand experience;
- Keep revising with specific feedback.
Early outputs were often stiff, full of stock phrases like "with the rapid development of AI" or "in summary." The real improvement didn't come from some magic prompt — it came from continuously feeding it samples and specific feedback.
AI can help take some of the pressure off starting from a blank page, but it isn't the one speaking for me. I'm still the one responsible for whatever gets published.
Second Team Member: The Information Organization Assistant
Businesses often have a lot of scattered information: data from different channels, day-to-day spreadsheets, customer records, and operational feedback. What managers actually want to see usually isn't the raw data, but results organized under consistent metric definitions.
I don't have the ability to independently build a database or a data platform on my own, so in this kind of scenario my role is mainly:
- Confirming with the user what results they actually need to see;
- Mapping out where the information comes from;
- Designing the aggregation and alert workflows;
- Using AI's help to generate or modify the implementation;
- Repeatedly testing whether the output matches business needs.
When it comes to data definitions, permissions, security, and complex analysis, data or technical staff still need to be involved. AI can cut down on repetitive organizing work, but it can't take over professional responsibility.
Third Team Member: The Operations Workflow Assistant
I also tried stringing together collection, filtering, copy generation, Feishu (Lark) review, and publishing for a Xiaohongshu (RED) multi-account operation.
My earlier talk described this as a stable, already-working automation system. To put it more accurately now: this project got through partial pipeline and feature testing, but as a whole it's still being iterated on — it isn't fully finished, and it shouldn't be taken as proof that I have independent development ability.
It still gave me some important lessons:
- Multiple accounts aren't simple copies of each other — each account's positioning and voice have to be defined separately;
- Human review must stay in place after content generation, not be removed;
- Platform rules and account safety matter more than "how automated" something is;
- One successful post doesn't mean the system can run stably long-term;
- Failure logs, retries, and manual takeover are all part of the workflow, not exceptions to it.
AI-generated code, produced by directing AI in natural language, lets a non-technical person see a prototype faster, but requirements, testing, and risk judgment can't be skipped.
Three Common Pitfalls in Managing AI
Handing every problem to AI
AI isn't suited to carry all the work. Tasks that need in-person communication, building trust, final decision-making, or high-risk judgment shouldn't be handed off to a model alone.
Assuming the first version will run stably
Any automation will run into problems with data quality, tool changes, permissions, network issues, and platform rules. The point of a first version is to surface problems, not to prove the thing is done.
Not having clear standards yourself
"Help me write a good article" or "help me build a good system" aren't actionable requests.
Who's the reader, what's the input, what's the output, which results are unacceptable, who signs off — the clearer these standards are, the more likely AI is to produce something useful.
Work That Shouldn't Be Handed to AI
At least three kinds of work call for extra caution:
- Work that depends on face-to-face communication and interpersonal trust;
- Work involving final decisions in medicine, law, finance, or business;
- Work where mistakes are costly but there's no reliable way to check for them.
AI can offer information, drafts, and suggestions, but it can't take on responsibility on a person's behalf.
A more reasonable division of labor is: AI handles the repetitive parts with clear rules, and humans handle judgment, relationships, creativity, and critical decision points.
How an Ordinary Person Can Get Started
Step one: don't start by thinking "I need to assemble an AI team." Start by finding the most annoying, most repetitive piece of work you have.
Step two: explain it the way you'd brief a new coworker — background, input, output, standards, and off-limits areas.
Step three: start with a small, testable version. Don't treat it as finished just because it runs; and don't write off the whole direction just because the first version is rough.
Step four: keep human review in place, and keep revising based on real errors.
With AI's help, one person can extend their own execution capacity — but this stays AI-assisted work, not a substitute for the person doing it. What actually determines the outcome is still whether you can clearly explain what you need, whether you can recognize risk, and whether you're willing to take responsibility for the final result.
That may be the meaning most worth holding on to in "a one-person AI team."