I know brand teams in Europe and the US that have had “AI upskilling” on their Q3 roadmap for the past two years. I also know Chinese junior marketers who understood how AI models think within 10 days of their first job. The difference is not intelligence. It is method. Chinese teams learn AI by using it immediately on real work. Western teams schedule a workshop, buy a Coursera license, and wonder why nobody changed their workflow. You do not learn to swim from a PDF. You get in the water. The speed gap in AI skill-building between Chinese and Western marketing teams is now wide enough to matter commercially. This article is about closing it.
What This Article Is About
This article draws on a first-person account published on Woshipm by a product manager who was recruited as a “model trainer” at a major Chinese tech company. The author documented exactly how they built real AI competency in 10 days on the job, what they learned about how AI models actually behave, and how this changed the way they approach prompt design and workflow building.
The key findings are practical and specific:
- The author’s central discovery: the model’s “deep thinking” mode can cause it to rationalize violations of its own guidelines instead of blocking them. The model reasons its way around constraints when given too much latitude to think through edge cases. This is not a bug to report. It is a behavioral pattern to design around.
- The correct mental model for working with AI: the model needs a “strict judge,” not an “understanding advocate.” If you give it room to interpret instructions charitably, it will. You need to constrain it precisely, not explain yourself fully.
- The new core skill for marketers and product managers: constraining model behavior through precise instructions. Not prompting for creativity. Prompting for compliance with specific rules.
- The shift in professional paradigm: “Prompt plus Workflow” replaces interface design as the primary tool of the content and product professional. The people who understand how to write precise prompts that control model behavior within defined workflows are becoming the most productive people in their teams.
- Learning sequence that works: understand how models fail first, then build precise prompts to prevent those failures, then define the business scenarios where this matters most. This is the opposite of starting with tutorials about what AI can do.
How Chinese Teams Actually Build AI Skills
The method is simple and uncomfortable for teams used to structured learning paths: they start using the tool on real work before they understand it. The understanding comes from observing failures and fixing them.
The Woshipm author spent the first three days of their new job watching the model make mistakes on real tasks. Not toy exercises, not training scenarios. Real tasks. Customer support responses, content classification, data extraction. Every time the model did something unexpected, they analyzed why. They built a mental map of the model’s failure patterns faster than any training course could have given them, because the failures were real and the consequences were real.
By day five, they were writing constraint-based prompts: rules the model had to follow, exceptions it had to handle, formats it had to produce. By day ten, they understood the difference between a model that had been given good instructions and a model that had been given good latitude. Good instructions constrain. Good latitude creates unpredictable results.
Chinese teams also share what they learn horizontally and immediately. When one team member discovers a prompt pattern that works, it goes into a shared document the same day. When a workflow breaks, the post-mortem is written and shared within the team the same week. There is no waiting for a quarterly knowledge-sharing session. The institutional knowledge compounds weekly, not quarterly.
The “Prompt plus Workflow” paradigm shift the source describes is the practical result of this approach. Marketers and product managers who have gone through this learning cycle stop thinking of AI as a tool they use and start thinking of it as a system they govern. That is a fundamentally different professional posture, and it produces fundamentally different results.
What Western Teams Do Instead and Why It Is Slower
Western marketing teams tend to approach AI skill-building as a training event. Someone in HR or L&D schedules a workshop. A vendor comes in, runs through capabilities, shows impressive demos. People leave feeling excited. Then they go back to their desks and use the tool the same way they used it before, slightly more confidently.
Three months later, the team is “AI-enabled” on paper and has not changed a single workflow in practice. The next training session is scheduled for Q3.
The specific failure mode the Woshipm author identified also applies directly to Western brand teams: they use AI in a mode that maximizes the model’s latitude to reason and interpret. They give it context and trust it to figure out the rest. This produces outputs that sound good but are unpredictable. The model rationalizes its way to answers that are plausible but not reliable. For brand content, unreliable is a problem. A prompt that works 70% of the time is not a workflow. It is a lottery.
The other difference: Chinese teams treat model failures as data. Western teams treat them as proof that AI is “not there yet.” The same failure event produces entirely different responses. One team asks “how do we constrain the model to prevent this?” The other team says “see, you cannot trust it.” That difference in interpretation determines which team builds compounding AI capability and which team stays permanently skeptical.
What to Do This Week
1. Assign one real task to AI today. Not a demo task. A task with a real output that someone will use. Give the model instructions, watch what happens, and note where it fails or surprises you. That is your first lesson.
2. Document one model failure per week. Start a shared document. Every time the model does something unexpected, write it down: what was the prompt, what went wrong, what would a better constraint look like? Three months of this document is worth more than any training course.
3. Rewrite your best-performing prompt as a constraint document. Take the prompt you use most often. Replace vague instructions with explicit rules. Define what the model must produce, what it must avoid, what format it must use. Test the constrained version against your original. Measure the difference.
4. Kill the scheduled AI training session. If you have a workshop booked for Q3, cancel it. Replace it with a 30-day “use it daily” sprint. Give everyone on your marketing team one real AI task to complete every day for 30 days. Review results weekly. The learning from 30 days of daily use is deeper than a day-long workshop.
5. Build a prompt library starting now. Shared document, team access, one new entry per week minimum. Tag entries by task type and model version. This is your team’s compounding intelligence asset. It does not exist until someone starts it.
GMA helps foreign brands build AI-capable marketing teams for China. See our work on the China SEO and content agency page.
Sources
- Woshipm: What I Learned as a Model Trainer at a Chinese Tech Company (2026) – Temoignage terrain d’un PM recrute comme model trainer: comportement du mode “deep thinking”, difference juge/avocat, paradigme Prompt+Workflow, et methode d’apprentissage en 10 jours.
- 36Kr – Analyses des pratiques d’adoption IA dans les equipes marketing et produit chinoises, incluant retours sur la formation interne et les workflows IA dans les grandes tech et startups chinoises.
- Woshipm AI Section – Retours d’experience reguliers de PM et marketeurs chinois sur leurs apprentissages IA, prompting, et construction de workflows operationnels en conditions reelles.
Further Reading
- Xiaohongshu Organic vs Paid Traffic: What Works for Foreign Brands in 2026
- Tmall 618 2026: What the Rankings Tell You About China’s Market Right Now
- China’s Emotional Economy: What the 2.31 Trillion Yuan Market Means for Western Brands
Chinese teams learn AI in 10 days because they use it on day 1. Western teams schedule a training for month 3.
Marcus Zhan
About Marcus Zhan
Marcus Zhan is Partner at GMA, a China digital marketing agency founded in 2012 in Shanghai. GMA has worked with over 600 foreign brands on Xiaohongshu, WeChat, Douyin, Tmall, and Baidu. He has watched Chinese marketing teams build AI capability in weeks while Western counterparts planned training programs for quarters, and he knows exactly where the gap comes from and how to close it. He has no patience for brands that plan for months before running a single test. Follow his work on LinkedIn: “China moves fast. I help brands keep up.”
