The Course Handbook

A large language model writes one token at a time by predicting what is most likely to come next. That is why it is fluent, and why it can be confidently wrong.

Models read in tokens (词元). Chinese uses roughly twice as many tokens per character as English does, which affects how much fits in the context window and what an API call costs.

Doubao for everyday tasks and voice, DeepSeek for reasoning, Kimi for long documents, Qwen for all-round work and images. Use more than one when the answer matters.

Many Chinese models are open-weight (开源): companies can download and run them on their own servers. This matters for data control and cost.

Give the AI a role (Act as…), explain the situation and what you need, then say please do… with the format, length and audience. It works in English and Chinese.

Give the AI a structure to fill, or ask it to reason step by step before answering. Both make complex answers more reliable.

Two or three examples of the output you want teach tone and format faster than any description.

Write the instructions in the language you think in, and name the output language and register you need, such as formal Chinese for a client (正式) or plain English for an overseas partner.

Use AI to widen your options and test your reasoning, not to make the decision for you.

Decision matrices and weighted scoring make trade-offs visible. Ask the AI to argue against your preferred option.

SWOT, scenario planning and going-global (出海) analysis work well with AI when you supply real data about your company.

Score AI work for accuracy, completeness, relevance and tone before you use it. You are responsible for what you send.

Chat answers one request. A workflow (工作流) runs fixed steps. An agent (智能体) decides its own steps towards a goal. Use the simplest one that does the job.

Coze for bots and workflows, Feishu Base and DingTalk for automations inside your work suite, Dify for IT teams that need self-hosting.

Let AI draft and sort; let people approve anything that goes to customers, moves money or touches personal data.

Score each agent on data sensitivity, permissions and what happens if it goes wrong, then start in draft-only mode.

Vibe coding means describing the tool you want in plain language, testing it, and asking for changes one at a time.

A one-page calculator, checklist or tracker in Kimi or Doubao teaches the method. Move to Trae or Tongyi Lingma for real projects.

Describe an app you like screen by screen and ask the AI to rebuild the parts you need.

Paste the exact error and what you expected. Ask for the cause before the fix.

Subject, style, lighting, composition and mood, in that order. Jimeng accepts English or Chinese.

Chinese image tools use on-screen settings: aspect ratio, reference image and number of results. Match the ratio to the platform: 3:4 for Xiaohongshu, 9:16 for Douyin.

Lock style, lighting and colour, and use a reference image to keep a product or character the same.

In Doubao, Qwen and Meitu, describe one change at a time: replace the background, change the light, remove an object.

Kimi, Qwen and Metaso can research a topic and cite sources. Always open the key sources before you rely on a number.

Kimi's PPT assistant and WPS AI build a draft deck from a brief. Spend your time on the story, then refine the slides.

Generate short clips in Seedance or Kling, then edit, subtitle and add music in Jianying (剪映).

For each department, keep a small library of tested prompts with the tool, time saved and who reviews the output.

Score your current AI use honestly. The aim is steady weekly practice, not knowing every tool.

Weeks 1–4 foundations, weeks 5–8 your role, weeks 9–12 a real project and teaching others.

Label AI-generated content you publish; protect personal information under PIPL; use approved services and your company's policy for confidential work.

A 30-minute session with one live demo and a one-page handout spreads AI skills faster than any memo.