AI Growth Hacking Weekly — EP#54: YingShiJuFeng’s Workflow, AI Viral Videos Deep Dive, Silicon Valley Token-Maxxing Arms Race, Growth Engineers’ Identity Crisis, DeepSeek Bestseller, and more
"I have found that Silicon Valley itself is starting to fall behind."
1/ I noticed that a significant portion of this newsletter’s subscribers come from countries and regions outside China; from private conversations, I’ve also found many employees and executives at global companies. So I’ve decided to start publishing an English version simultaneously, starting from this issue.
To avoid disturbing everyone, the Chinese version will continue as usual — sent via Email and Substack App. The new English version will initially only be published via Substack App, and won’t be sent via email for now. This way, existing subscribers won’t immediately receive two emails (the same content in Chinese and English), adding to their information burden.
Going forward, I’ll find ways to transition gradually. For example, if you only need the English version email, just send me a private message or reply, and I’ll tag you accordingly, so we can push emails based on tags going forward. Want to give it a try? Or if you have better suggestions, please don’t hesitate to share.
Today’s first English version will be published on Substack right after this Chinese version goes out.
2/ Friends in China are currently on May Day holiday. Among them, some people don’t like crowds, deliberately avoiding peak tourist flows, and choose to stay home for a “learning-focused holiday” — cramming up on AI knowledge.
For this reason, I’ve specially set the two popular AI video courses I created to a limited-time (May 1-7) special price: 299 yuan → 99 yuan to support your holiday learning plan.
Click to buy: How I Built a 100X Knowledge Extraction System with AI
Click to buy: How I Built My Personal AI Jarvis Assistant
The first course teaches you how to use Obsidian + AI to build your own information system; the second teaches you how to use AI tools to build your own agent (similar to OpenClaw / Hermes, but completely something you can build and control yourself — your own secret weapon).
Both courses were recorded a while back. I know the AI world changes daily, but these two are still worth buying (during the sale) because I followed a long-term perspective from the start: all my tool choices and technical architecture are still applicable today.
When I advocated for Obsidian, it wasn’t this popular yet. Now Obsidian has 5 million monthly active users and is taking significant market share from Notion as the go-to AI knowledge worker tool.
When I taught everyone to build their own Agents, OpenClaw and Hermes experienced explosive growth then cooled down, while my system, built on a clear architecture, has been continuously absorbing their essence and becoming more stable and powerful.
In fact, the upgraded version I currently use (not publicly available) is still based on what I learned from these two courses, iterated over time. So if you’re just starting to build your own AI knowledge base or develop your own Agent, you can learn from my courses. AI experts can bypass this — it might be too basic for you.
Wish everyone learns by applying, solves valuable real problems, and truly lets AI reduce your costs, increase efficiency, and earn more money.
OK, here’s the content this week — Enjoy!
▪️CASE Studies
How Does YingShiJuFeng Actually Work in 2026?
via YingShiJuFeng
YingShiJuFeng’s team recently released a self-documentary saying that from 2024 to 2026, they’ve deeply integrated AI into their creative workflow.
The biggest changes are in post-production. They differentiate between two lines: the short video team uses AI to lead creation, while the main channel uses AI-assisted — using AI voice models to replace real recordings for internal rhythm references and client review; using AI video generation to supplement shooting and special effects, like dream sequences; also using open-source tools for AI rotoscoping.
But where AI truly delivers value is at the front end. For topic selection, they deployed a Feishu bot to automatically gather information and generate covers, using AI as a creative brainstorming card draw. The planning phase is even more disruptive: previously it was text-to-image, now AI generates reference videos first to eliminate information gaps and communication costs. Shooting changes are minimal, but equipment management achieved automated inventory in/out, with underwater recording systems being an innovation highlight.
They also built a data hub, using AI to analyze the relationship between disc inclusion metrics and retention rates; they hold regular AI sharing sessions, encouraging everyone to use programming tools to build internal tools.
One-sentence summary: AI’s value lies in guiding direction in the early stage and improving efficiency in the later stage, but ultimately implementation still requires human execution.
Analyzing 100+ AI Viral Videos: 6 Key Observations
via Yun Feiyang 1993
After analyzing 105 AI videos with 1 million+ likes on Douyin, the author found that transition and costume change is the biggest category, accounting for over 30%. Blogger “Jian Peng” earned 13.65 million likes on an 11-second AI video — the highest viral generation rate among creators.
But pure AI short films are also starting to go viral. “Paper Phone” achieved 230 million total exposure through emotional resonance, proving AI video has entered the mainstream narrative.
Interesting conclusions at this stage:
Real person + AI combination is still the mainstream, with significantly higher viral rates than pure AIGC videos. Funny bloggers like “Hou Lüluo” and sketch bloggers achieved good results through AI-assisted brainstorming.
Smaller accounts are actually more likely to go viral. For example, “Zhi Jing Yan” could produce viral hits because AI lowered the barrier to creation. But creators need sustained output ability — going viral isn’t available to everyone.
What ultimately determines success is emotion, creativity, and aesthetics. “Jian Peng” used AI to evoke nostalgia; “Not This Universe” used AI to “resurrect” loved ones — what makes netizens stop and like is the story and emotion behind the AI, not the technology itself.
A DeepSeek Operations Book Sold 23 Million Yuan — The AI Book Publishing Gold Rush
via Yang Jing
Famous publisher “Qiuye Dushu” caught the window when DeepSeek exploded in early 2025 but ordinary people didn’t know how to use it. He used AI to reconstruct the book-writing process, compressing what was originally a 5-6 month publishing cycle to 1 month. The final book “AI Era Survival Manual: Zero Basics to Mastering DeepSeek” sold 400,000 copies, with total retail value exceeding 23 million yuan.
His business model isn’t complicated: books are low-price traffic generators, training camps are the profit center. The real profit comes from back-end conversion — 50% of users are workplace professionals, 40% are stay-at-home moms, all groups willing to pay for “hand-in-hand teaching.”
But Qiuye himself is also anxious. AI iteration speed is so fast that he “can’t understand or see clearly,” and employees also feel “the heart is willing but the body is not.” He’s caught in a paradox: teaching people to use AI, but being pushed by AI himself.
Looking back at his path: starting from a casual PPT training invitation in 2009, he’s hit every knowledge-payment wave over seventeen years — from Weibo to WeChat Official Accounts to short video. After announcing All IN AI in 2023, he’s published nearly 40 AI teaching books with total sales exceeding 3 million copies.
The wealth creation opportunity always belongs to those who can quickly turn tech anxiety into consumer demand and use industrial processes to mass-produce content. But the window for this dividend is narrowing.
▪️OPINION
Full Team Token-Maxxing: An Arms Race No One Can Stop
via MengXing
Strongly recommend this piece — it’s been a while since I’ve seen such an information-dense and authentic article from a first-line VC practitioner. The author offers many insights:
80% of YC’s latest batch are vertical agents, but companies selected 5 months ago have already lost investment value due to AI iteration speed. YC itself has changed from a direction leader to a lagging indicator.
Meta’s entire team uses competitor Claude Code to write code, with code security thrown by the wayside. Internally, they even have token consumption leaderboards — bottom performers may face layoffs. At AI-native startups, one engineer’s annual token budget is approaching their salary.
The core subtext: everyone is betting on rushing speed first — total costs haven’t decreased, just shifted from human labor costs to token costs. Silicon Valley itself can’t keep up with its own rhythm anymore.
There’s no exit from this race. The only choice is to continue doubling down.
Building a 70% Day-One Retention Personal Agent: Reflections
via Eva433
The author built a personal assistant product called Poke, achieving 70% day-one retention and 30% monthly retention. He reflects on everything from getting inspired to developing it himself.
Poke’s biggest inspiration was making AI chat like a friend — proactively breaking the ice. But the downside was the positioning was too broad — users didn’t know what this thing could actually do or what they’d be willing to pay for.
He then cut into the “proactive monitoring and reminders” niche. Drawing inspiration from the real-person supervision service on Xianyu, he discovered users truly need the value of “being reminded, being seen, being pushed” — not chatting, but having someone watching over you.
Based on this insight, he quickly built an Agent framework with a heartbeat mechanism, adopting an aggressive pricing strategy (199 yuan/month) to filter for the most pained users and validate the MVP.
For the 0-1 stage products, his experience is: don’t try to satisfy everyone’s all needs — must focus; launching first then iterating beats pursuing perfection, because actual user feedback forces fixing real pain points.
One more reminder: early testing should exclude product managers and investors — their needs don’t represent real users.
An 80-Year-Old Investment Legend’s 10,000-Word AI Insights
via Howard Marks
Howard Marks, co-founder of Oaktree Capital, is nearly 80 this year. He published a supplementary memo called “AI Surging Forward,” using his fifty-year investment framework to dissect this technological revolution.
His judgment refreshes perception: AI is absolutely not a search engine, but an intelligent system trained to reason and synthesize data. Development has moved from chat interaction and tool usage into a new stage of autonomous agents, upgrading from productivity tool to labor replacement, spreading far faster than computers and the internet.
He both affirms AI’s real value and underestimated potential (limitations come from users, not models), while warning about hallucinations, reliability issues, and impacts on the job market.
At the investment level, he points out AI has advantages in rational analysis, yet struggles to match human judgment and intuition in entirely new scenarios.
Facing market frenzy, his advice is: don’t blindly go all-in, don’t completely exit — embrace this era-defining transformation with moderate positions and selected targets.
Marks isn’t simply commenting on AI — he brings his own methodology into this new territory, using a stable framework as a tool for questioning and calibration. This learning ability and curiosity is exactly what makes him admirable.
When AI Can Run the Entire Loop, What’s Left for Growth Engineers? — Lemon Latte EP01
via Bu Gudu De Er Xiang Bo
In podcast “Lemon Latte” EP01, two growth engineers discuss their respective practice paths.
Xmind’s growth team lead Zhang Xiaoji’s core is “collaborative humans” — connecting AI capabilities to team workflows, like using CLI tools so operations colleagues can publish small tool pages with one click, achieving 100-200 pages launched per week, with 95% done by AI.
The author emphasizes the core is “collaborative agent” — in a scenario where one person handles growth engineering, orchestrating an agent cluster to chase trends and build automated loops. For example, when Google released Nano Banana 2 that night, he prepared prompts and frameworks in advance, letting agents check for model launch every 30 minutes. Once launched, they immediately launched in parallel the entire process: batch image generation, deploying comparison sites, writing SEO blog posts, generating KOL promotion copy, etc.
What a traditional ten-person team produces in a week, one person can now roll out in an afternoon. Using quantity to trade for probability has become the new growth posture.
Neither path is superior — it depends on team stage: multi-person teams are better suited to Zhang Xiaoji’s path, while individuals or small teams are better suited to the author’s approach.
The value of growth engineers no longer lies in executing themselves, but in designing AI collaboration systems for decision-making and review.
Quick Reads:
Some Uninstalled Within 10 Days, Some Already Making 5,000 Yuan Monthly! The Hangzhou OpenClaw Trend — How Many Are Still Using It? interviews four Hangzhou office workers still “raising shrimp” (using OpenClaw). They deeply embed AI tools in specific business pain points. R&D engineer Yang Xianjun automatically pushes morning reports daily, can convert group chat records into requirement documents, and spots million-level risk terms in contracts during review; cross-border e-commerce operator Chen Ying used OpenClaw to complete detailed analysis of 500 competitor SKUs in one day. For those who can wield it, OpenClaw’s monthly cost of a few hundred yuan can exchange for over 15% R&D efficiency improvement or multiple content output improvements.
Six Months of Entrepreneurial Coding: Smooth Sailing, Rough Patches, and What I’ve Never Figured Out is a six-month retrospective from a byte-hopping entrepreneur named Meng Jian. He cracked the overseas expansion loop, deployed OpenClaw for traffic, exceeded expectations with the Navigation Plan enrollment, and has good cash flow. But Resend account was banned, Creem was banned for data leaks, OpenClaw book sales fell far short of expectations, YouTube and Reddit accounts were restricted for posting external links. He cites Li Shanglong and Robert Kiyosaki’s ESBI quadrant theory, positioning his migration from E/S quadrants to B/I — core is shifting from pursuing security to pursuing freedom.
After Years of Reporting AI, I Feel More Like a Fraud... uses Anthropic’s Opus 4.7 as an example — benchmark scores seem to improve, but actual user experience has regressed, with developers even switching back to old versions. Baudrillard’s “simulacra and simulation” theory is extremely apt here: AI benchmark tests and hype have replaced actual experience. Hype has reshaped reality — the model is no longer the product; the release is the real product.
Generation 10s, Starting to Use AI to Compile Their Own Huang Gang Secret Papers describes how the new generation of primary and secondary students are using AI as a learning tool, proactively customizing their own learning content. 14-year-old Shenzhen middle school student Qianqian, preparing for the physics Olympiad elimination exam, built a physics learning website with his father using AI; 13-year-old Lü Sitong developed a 24-hour online “Frog Foreign Teacher” using AI to strengthen oral English. This group of digital natives is turning AI into “the most handy clay,” quietly upgrading their problem-decomposition ability, structured expression ability, and confidence.
A Girl from a Non-Top University Used Free AI to Get into Peking University Protagonist Chen Yuxin, from Anhui, failed the college entrance exam and entered a second-tier college, determined to apply to Peking University’s Social Work major. In an isolated environment, she initially relied entirely on handwriting nearly 300,000 characters of notes. The turning point was systematically using free AI tools: using AI to organize subject frameworks, deconstruct exam question logic, simulate interview answers, generate personalized learning plans. AI didn’t replace her effort — it served as an “all-day personal tutor” and “information filter.”
One Afternoon, One Sentence — Codex Helped Me Develop a Complete Game Author “Guizang 2079” used just one sentence to let Codex call built-in GPT-Image 2.0 to generate materials, and Codex automatically planned a complete character asset pipeline. When the asset site restricted downloads, Codex itself found and integrated a free icon library. Throughout the entire development process, the author was only responsible for the goal — “I want a card game similar to Slay the Spire” — and controlling the final visual and experience aesthetic standards. The fundamental difference between Codex and Claude Code lies in its built-in browser, image model, and do-or-die execution drive. The combination of all three fundamentally transforms the developer role from “configuring tools” to “setting goals and making aesthetic decisions.”
Tutorials & Resources for Individuals:
My Most Used AI Products in 2026 is AI self-media personality “数字生命卡兹克” publishing a product list before the May Day holiday, covering both overseas and domestic alternatives. For knowledge Q&A, GPT-5.5 has extremely low hallucination, and domestically you can use Doubao; for content creation and knowledge management, Claude Opus 4.6 is in a league of its own, with DeepSeek V4 Pro recommended domestically; for data analysis, use Codex + GPT-5.5, or Claude Code + GLM-5.1 domestically; for image design, GPT-image-2 is in a league of its own, and domestically use YiDream Seedream-5.0-lite; for video generation, Seedance 2.0 emphasizes motion texture, KeLing 3.0 focuses on cinematic texture and native 4K. Covering 14 scenarios, emphasizing practical operation and alternatives.
Open-Sourcing a PPT Skill: 10 Years of Design Experience Compressed: This tool called guizang-ppt-skill isn’t a simple template, but a system defining the human-machine collaboration interface. The real value lies in three things: First, a pre-clarification process — AI proactively asks six key questions about audience, duration, materials, etc., aligns before writing code, intercepting 80% of rework; second, standardized image naming and storage rules, achieving lossless image replacement through same-name overwrite; third, translating magazine industry layout language into rigid rules, such as three-level font hierarchy, color discipline, and maintaining rhythm through fixed grids alternating between hero pages and non-hero pages.
Best Video Creation Workflow: Image2 + Seedance 2.0, Topview One-Click Loop | Cross-Border E-commerce Version: Image2 handles pixel-precise static image control, Seedance 2.0 serves as the storyboard execution engine, driving camera movement and character actions. Key insight in prompt design: Image2 needs to use physical facts like “soft afternoon light hitting from the left at 45 degrees” instead of empty words like “minimalist” or “cinematic”; Seedance 2.0’s prompts should be extremely short, prioritizing description of subject action and camera movement. Five-layer prompt framework and three iron rules: duration must match instruction density, one shot does one thing, reference images stack in priority order of character-face-scene.
D49 | Today We Open-Sourced an AutoResearch Tool discovered AI execution has “execution out-of-control syndrome” — tasks break, status goes blind, results fake, direction drifts. So they developed Thoth. Solution: using callback hooks and watchdogs to keep sessions alive; using a human-facing dashboard to present progress; task verification must pass through mechanized script arbitration, never AI self-judgment. Core insight: transforming AI execution “trust” into verifiable “mechanical contracts” is what truly makes AI reliable.
Turning Paper Reading into an Assembly Line: Four Quant Research Skills from arxiv Search to Event-Driven Strategy breaks quant paper research into four Claude Code skills, forming an automated pipeline from search to strategy implementation. Skill 1 searches and downloads PDFs from arxiv; Skill 2 extracts PDFs into structured research notes and machine-readable indicator JSONs; Skill 3 combines scheme design, backtesting, and benchmark comparison in three stages, requiring plan.md writing first waiting for user confirmation; Skill 4 translates verified strategies into event-driven files. The author emphasizes: if reproduction is too clean, be wary — must honestly mark the source of differences.