The 13-lesson AI self-media course released by dontbesilent (Old D) in 2026-05 systematically teaches users in mainland China who do not use scientific internet access and have no coding background the complete production chain for doing self-media with AI.
Core Stance:
- I will not teach you scientific internet access,All SOPs work without scientific internet access(using DeepSeek + domestic wrappers + Openrouter)
- No technical foundation required,Claude Code is designed for direct use by non-programmers
- Do not pursue "perfectly removing the AI flavor",Keep your own style instead of eliminating AI flavor
1. Three-Layer Prompt Architecture (core mindset running through the entire course)
我写"元提示词"(Meta Prompt,我用键盘敲)
↓ 输入给 Gemini / Claude / DeepSeek 等
AI 输出"系统提示词"(System Prompt,可能 2000-3000 字)
↓ 复制到智能体平台
变成"智能体"(Agent)
↓ 接受用户输入
用户输入"用户提示词"(User Prompt)
↓
智能体输出回复
Key Insight:
- Meta prompt = a prompt used to produce prompts. It is like a Russian doll, but solves the problem that "humans are not good at writing long prompts"
- Do not write system prompts directly: let AI write them; you only write the meta prompt
- Domestic first choice for writing meta prompts is Gemini 2.5 Pro(long context + strong reasoning);Use DeepSeek R1 to drive agents(usable in China + if a prompt works on R1, it is generally OK across the board)
2. L01 Large Model Selection (domestic and overseas paths)
Overseas model selection preferences
| Model | Recommended Scenarios | Notes |
|---|---|---|
| GPT-4o / 4.5 | Fallback general-purpose option, optional | Official website quality degradation issue (using a VPN will secretly downgrade you to mini/3.5) |
| Claude 3.7 Sonnet | Advantages in quick Q&A / partial writing tasks | Easy to get account banned |
| Gemini 2.5 Pro | Complex tasks / long context / reasoning | Old D's personal daily driver |
| Grok 3 | Another option for writing | Deepsearch feature |
| GPT o3 | Image/hidden information reasoning | Private detective-level recognition |
Reasoning models vs general models
| Dimension | Reasoning (R1 / o1 / o3 / Gemini 2.5 Pro) | General (V3 / 4o / Claude 3.7) |
|---|---|---|
| Response speed | Slow (displays think process) | Fast |
| Compute cost | High | Low |
| Suitable tasks | Mathematical derivation / logical analysis / code generation / complex planning | Text generation / creative writing / multi-turn dialogue |
| Startup latency | Thinks for a long time | Replies instantly |
Old D's preference: use Gemini 2.5 Pro for complex tasks, use Claude 3.7 when speed is needed.
Domestic / no scientific internet access first choice
- DeepSeek R1 + V3 — usable in China + performance in the first international tier
- Monica wrapper — accesses all overseas models from within China, annual fee ~1000 yuan, compliant and usable
- Openrouter — US dollar top-up, supports small amounts (starting at ¥40), first choice for developers
- Tencent Yuanqi(yuanqi.tencent.com) — domestic agent platform, usable by scanning QR code without login, model can be switched to R1
Model selection ≠ the cure for insufficient prompts
"Many people want to switch to Claude / Gemini when GPT doesn't write well. But they ignore that the core is the prompt itself. Improving prompts is more important than switching models."
3. L02 The Correct Method for AI Copywriting Imitation
Root cause of 90% of imitation failures
It is not that the model is bad, but that you did not define what "imitation" is.
Key elements of imitation:
- Vocabulary and wording?
- Length? (exact word count / roughly similar)
- Tone?
- Style?
- Keep punctuation?
- Number / position of emojis?
4-step imitation SOP
1. 准备提示词模板(Markdown 结构化)
# 你的任务
先分析参考文案,反向向用户提问列表,确认所有关键要素后再仿写
# 参考文案一
<粘贴文案 1>
# 参考文案二
<粘贴文案 2>
2. AI 接收后不立即仿写,而是发 8 个左右问题(字数/结构/语调/风格/用词偏好/特殊要求/EMOJI/长度)
3. 用户逐一回答 → AI 仿写
4. 不满意 → 用问题修订(列出 3 条不满意 + 修正)→ 再生成
Key points:
- Use Markdown
#layering to prevent AI from treating multiple reference copies as one - One reference case is enough (if the case itself contains all features); otherwise you need multiple mutually consistent cases
- Do not give messy, inconsistent cases — AI cannot find common ground
4. L03 The Essence of Knowledge Bases and Agents
RAG Architecture (Retrieval-Augmented Generation)
Most common misconception:"Knowledge bases make large models stronger."
Truth: The model does not change. RAG only modifies your prompt input.
工作顺序(以ima知识库为例):
1. 用户输入提问("做生意第一步...")
2. RAG 用向量检索去文档库找相关资料(找出 27 篇)
3. 把 [27 篇资料 + 用户问题] 一起灌给大模型
4. 大模型基于 28 个信息源综合输出
Why you can get good answers even with rough questions: It is not that the model becomes smarter, but that the prompt is automatically enhanced.
Knowledge bases are actually agents in disguise
Evidence:
- When you create an ima knowledge base, there is actually a system prompt inside (you just cannot see it)
- When a question goes beyond the knowledge base scope, it actively refuses to answer(indicating a built-in instruction of "if the document does not contain it, refuse to answer")
Perplexity uses the same architecture:
- Perplexity = knowledge base + the internet as the "knowledge source"
- ima knowledge base = documents you upload as the "knowledge source"
- Their underlying logic is exactly the same, only the Source is different
Practical: Build a personal AI avatar
1. 收集你的内容:推文 + 直播文稿 + 微信群问答 + 付费资料 + 写过的书
2. 上传到ima知识库(腾讯)或元气
3. 设定系统提示词:"你是 X 的 AI 分身,基于他的内容回答..."
4. 问问题 → 比直接问大模型多 90% 个性 + 风格
Old D's own AI avatar already has 1000+ users.
5. L04 Using Knowledge Bases to Assist Talking-Head Videos
Starting point for business judgment
4 types of large model outputs: text / image / video / music + MCP.
Old D's judgment:
- Only text meets commercial standards (after 2025 images and videos also work, but costs are not necessarily lower than real people)
- Text itself does not make money,monetization is easiest when combined with a profitable physical business(custom suit reports + five elements / Bazi fortune reports)
- Self-media talking-head videos:Let AI write the script, the person appears on camera + edits + adds subtitles themselves
Three elements of a script
| Element | Handling |
|---|---|
| Viewpoint | Let knowledge base + AI produce comprehensively (may be counter-intuitive / sharp) |
| Format | Decided by the user themselves |
| Expression style | Keep personal style, do not remove AI flavor |
"Do not pursue eliminating AI flavor — make the script sound like you, so that hearing the voice feels like it is you"
6. L06 AI Drama Short Videos (text-to-image / image-to-video / lip-sync)
Workflow
Text-to-image → select satisfactory character → image-to-video (Jimeng/Vidu) → lip-sync (voice finalized).
Key critical tools
- Jimeng — text-to-image + image-to-video + character fusion
- Vidu — supplementary video generation
- HeyGen / Tongyi Wanxiang — lip-sync / digital human (L10)
(For specific steps refer to Caviar AI Comic Drama Production Course / Section 1 Live Teaching Complete Methodology for a more complete comic drama production process)
7. L07 Creating AI Workflows with N8N
Core: Turn multi-Agent chains into workflows
- N8N = open-source Zapier alternative + more AI-friendly integration
- MCP Trigger = Expose N8N workflows as MCP servers to other Agents
- toolWorkflow = N8N workflows are called as tools by external Agents
(See details in newtype Miscellaneous Practical Combat: llms.txt Protocol + CrewAI Obsidian Project + n8n MCP Workflow N8N MCP integration section)
8. L08 Efficiently Comparing Outputs from Different Large Models
Old D's comparison method
1. 同一个提示词 同时丢到 3-4 个模型(GPT / Claude / Gemini / DeepSeek)
2. 横向比较输出 — 看谁的风格 / 准确性 / 篇幅 / 创意更符合需求
3. 复杂任务做"投票":3 个模型同意的更可信
Tools:
- OpenRouter Playground — run the same prompt across multiple models for comparison
- Monica wrapper — switch between different models with one click
9. L09 Using Claude Code with Zero Technical Foundation
Core: usable without a paid subscription and without knowing code
Installation steps:
- Download an AI-powered command-line tool(to replace the native Terminal) — when errors occur you can ask directly, and AI helps you solve them
- Go to the Claude Code official website, copy the one-line installation command → paste it
- When errors occur, the AI command line gives prompts → click install
- Use Openrouter instead of Claude official membership
- Use a third-party tutorial to connect your Openrouter Key to Claude Code
- Launch Claude Code → chat with it directly in Chinese
Practical case: transcribe two audio files:
我 → "这个文件夹有 2 个音频,用 ElevenLabs API 转成文本"
Claude Code → 自动写 Python 脚本调 API → 报错自动改 → 生成 TXT
我 → "再把 TXT 转 Word"
Claude Code → 又自动写代码 → 完成
Key conventions:
- Use
Don't ask againmode to let Claude Code automatically approve permission requests - API Key exposure note: revoke immediately after demo
Difference between Agent and large model
| Dimension | Large model | Agent |
|---|---|---|
| Capability | Only generates text | Can read/write files / run commands / call APIs |
| Interaction | One question one answer | Autonomously executes multiple consecutive steps |
| Domestic "intelligent agent" translation | (not actually a real agent) | Claude Code is a real agent |
10. L10 LobeChat Blogger Imitation + Conversation Branching + Artifacts
Use LobeChat to imitate any blogger's title style
4 steps:
- Find the blogger's hit title data (cover title + text title + like count)
- Use Markdown structured prompts → let AI analyze the underlying logic (attention economy / human nature / communication studies)
- Based on the analysis, let AI write a system prompt
- Make the system prompt + Claude 3.7 Sonnet into an Agent
Artifacts whiteboard feature
- By default large models put long outputs in the chat box → the conversation becomes messy
- Turn on Artifacts (whiteboard) → long outputs go to a separate page → chat area stays clean
Conversation Branch — killer feature
对话: "请帮我分析陈浪浪文案"
↓ AI 给出分析
分支 A: 基于分析写系统提示词
分支 B: 基于分析写公众号文章
分支 C: 基于分析写微博
Each branch is independent,you do not need to build an agent first to maintain context, just split the conversation directly.
Old D's comment: Many people ask "How do I turn my analysis into an agent to preserve context" — you do not need to make an agent, LobeChat's branching solves it.
LobeChat configuration
- Client:LobeHub / Web version:LobeChat
- Payment: official website membership is available;or bind an Openrouter Key(cheap)
- Open source, full code visible on GitHub
11. L11 Making Digital Humans with HeyGen
Complete workflow (5-tool combination)
1. 豆包生图(白底图 ★ 关键)
→ 提示词:让豆包反推一段 prompt 给你
→ 生成数字人图片(白底,方便后续抠人)
2. 真人录视频(你自己拍)
3. 阿里通义万相 wan2.2 animate(角色替换)
→ 视频 → 数字人替代脸
→ 但声音还是你的男声 ← 问题
4. ElevenLabs Voice Changer
→ 用女声换音色(支持 29 国语言)
→ 选中文女声(如安娜苏)
→ 输出 MP3
5. 剪映合成
→ 音频替换 + 响度统一 + 字幕
6. (可选)HeyGen 翻译视频
→ 一键翻译成日 / 法 / 韩 / 西班牙 / 俄等 30 国语言
→ 抖音覆盖海外市场
Key conventions
| Pitfall | Solution |
|---|---|
| Video images must be white-background images | Otherwise background recognition will be messed up (walls become distorted) |
| Tongyi Wanxiang requires a paid membership | Standard version has queueing and lag |
| Different tools have inconsistent durations | Align to the shortest one and cut in CapCut |
| Select target language for HeyGen translation | If not selected it auto-detects but is error-prone |
Practical returns
A video Old D posted on Douyin yesterday got nearly 2 million views; after the digital human covered 30 languages single-video traffic doubled.
12. L12 One-Click Fully Managed AI Auto-Publishing to Xiaohongshu
Workflow (based on RAP)
RAP = rap.dev, AI-powered command line + one-click template distribution.
1. 在 RAP 输入:"帮我把这条推文发布到小红书"+ 链接
2. RAP 调用:
a. Playwright 操控无头浏览器截图(确保 3:4 比例)
b. 字数多 → 左右填白边(Python Pillow)
c. 字数少 → 拉远拍 → 裁剪
3. Claude 4.5 Sonnet 给小红书写标题(自动 ≤ 20 字 / 从原文找金句 / 不要 AI 写新词)
4. 小红书 MCP server 自动登录发布
- 截图 + 标题 + 正文 + tag = 一键完成
5. 还可以定时发布("1 小时后发")
Key technologies
| Component | Function |
|---|---|
| rap.dev | AI command line + one-click template distribution |
| Playwright | Controls headless browser to take screenshots |
| Python Pillow | Image filling / cropping |
| Xiaohongshu MCP server | Auto login + publishing |
| Claude 4.5 Sonnet | Write titles (stable at 80 points, no need for manual selection of 90-point ones) |
Key conventions
- Templates are released by Old D on GitHub; paste the link and RAP automatically installs all dependencies
- Without a membership, adding 5 US dollars on Openrouter is enough
- Never post API keys online / on WeChat / in documents
13. L13 AI Interpretation of Business Cases
Business case workflow
1. 选商业案例(瑞幸 / 拼多多 / 名创优品...)
2. 喂给 AI 多个维度提示词:
- 商业模式 5 要素
- 增长引擎拆解
- 财务模型推演
- 竞争壁垒分析
3. AI 出深度报告 → 你审阅 → 修订
4. 改写成自媒体文稿
Core mindset:Treat AI as a deep researcher, and be the editor-in-chief yourself.
14. L14 Claude Skill Engineered Prompt Management
Pain point
Well-written system prompts are scattered across various platforms:
- One version on LobeChat
- One version on Yuanqi
- One version on ChatGPT
- Different models have different effects
Result: good prompts cannot be reused, versions run wild.
Claude Skill solution
Skill = a "prompt engineering product" that can be called by Claude Code / other Agents:
SKILL.md= metadata + frontmatter + name + description + trigger conditionsreferences/*.md= long documents are not written in the main SKILL file, but are automatically loaded by Claude when the Skill is triggeredscripts/= supporting tools (such as scripts / data files)
Advantages:
- Centralized management of scattered prompts(version control available with Git)
- Automatic trigger conditions(similar to function signatures)
- Cross-platform reuse(any Skill-compatible Agent can call it)
Practical process
- Organize the prompts you have created into SKILL.md
- Use Claude 4.5 to organize frontmatter (name / description trigger conditions)
- Split long references into references/*.md
- Test trigger conditions
- Git commit = version management + team sharing
(See newtype Skill System in Practice (Super Analyst v1/v2 + Super Writer) for the Super Analyst Pro 2.0 Skill bundle design)
Full Course Methodology Overview Diagram
0. 提示词三层心法(贯穿全课)
元提示词 → 系统提示词 → 智能体 → 用户提示词
1. 模型选用基础(L01)
国内:DeepSeek + Monica + 元气
国外:Gemini 2.5 Pro + Claude 3.7 + GPT-4o
2. 文字仿写 SOP(L02)
Markdown 结构化提示词 + 反向提问 + 迭代修订
3. RAG / 知识库本质(L03)
不是模型变强,是提示词被自动加强
ima知识库 / 元气 / Perplexity 都用 RAG
4. AI 自媒体口播(L04)
AI 写稿 + 真人出镜 + 保留个人风格
5-7. AI 视频 / 工作流 / 多模型对比(L05/L06/L07)
即梦 + Vidu + N8N + OpenRouter Playground
8. Claude Code 零代码 Agent(L08)
Openrouter + AI 命令行 + 不开会员
9. LobeChat 高级技巧(L09)
Artifacts 白板 + 对话分支 + 模仿博主
10. HeyGen 数字人(L10)
豆包白底图 + 通义万相 wan2.2 animate + ElevenLabs + 剪映 + HeyGen 翻译
11. 一键发布小红书(L11)
rap.dev + Playwright + Claude 4.5 + 小红书 MCP
12. AI 解读商业案例(L12)
多维度提示词 + 自己做总编辑
13. Claude Skill 工程化(L13)
SKILL.md + references/ + Git 版本管理
Link to existing wiki
- Caviar AI Comic Drama Production Course / Lesson 1 Live Teaching Complete Methodology — More detailed video production workflow for the Caviar Comic Drama Course (storyboard / asset library / Jimeng / Vidu)
- newtype Skill System in Practice (Super Analyst v1/v2 + Super Writer) — Super Analyst three-piece Skill design, deeper productization of L13
- newtype Miscellaneous Practice: llms.txt Protocol + CrewAI Obsidian Project + n8n MCP Workflow — Complete N8N MCP server integration (technical details of L06)
- newtype Prompt Template Library (4 production-grade prompts + 1 CLAUDE.md sample) — Prompt template collection (YouTube Gemini v2 / Social Media breakdown)
- newtype Tool Stack: Oh My OpenCode Customization + newtype CLI Multi-Agent Team — Oh My OpenCode multi-agent customization (higher-order L08+L13 combination)
- Wang Kai Collection / 2024-2026 In-Depth Reading of Cutting-Edge AI Papers and Reports — Lee Boonstra Prompt Engineering provides a tactical framework
Quick Reference for Key Tools
| Tool | Use Case | Notes |
|---|---|---|
| Gemini 2.5 Pro | Writing meta-prompts | Ultra-long context |
| DeepSeek R1 | Driving agents | Available in mainland China |
| Monica | Overseas model wrapper | Annual fee ¥1000 |
| Openrouter | US dollar top-up for multi-model API | Starting at ¥40 |
| Tencent Yuanqi | Domestic agent platform | Use by scanning a QR code without logging in |
| ima Knowledge Base | Domestic knowledge base platform | Built-in RAG |
| LobeChat / LobeHub | Multi-model client | Open source + conversation branching + Artifacts |
| Claude Code | No-code Agent | Connected via OpenRouter |
| rap.dev | AI command line + template distribution | One-click install for Xiaohongshu publishing |
| Jimeng / Vidu | AI video | See Caviar Comic Drama Course for details |
| Alibaba Tongyi Wanxiang wan2.2 animate | Character replacement | Must use white background image |
| ElevenLabs Voice Changer | Voice timbre replacement + 29 languages | Monthly membership |
| HeyGen | One-click video translation into 30 languages | Monthly membership |
| Xiaohongshu MCP server | Automatic publishing | Open-source version released by Old D on GitHub |
| N8N | AI workflow orchestration | Open source + bidirectional MCP exposure |
| Jianying | Video editing | Domestic + free |
| Doubao | Chinese text-to-image + word-count reasoning | Top choice in China |
| Perplexity | AI search engine | Timeliness issues (Google alternative) |