The Launch of *AI Leadership*: Reflections on Change, Connection, and Responsibility

AI’s Constant and Changing

My new book AI Leadership, co-authored with Xiang Yang, is now published and available on JD.com and other major platforms. The book is a grounded exploration of AI’s foundational logic, written to sharpen cognitive clarity.

As I revisited old notes and drafts for this launch, one thing stood out: the pace of terminology change in AI is dizzying.

By late 2023, we were debating whether ChatGPT would replace search engines. Then came prompt engineering, knowledge bases, reasoning models, and multimodality. Next: Agents, workflows, MCP, Skill, digital employees.

Over the past two years, we’ve also begun actively studying and practicing GEO—thinking deliberately about how our brand, content, and expertise are discovered, interpreted, and cited by AI systems.

Just as you master one concept, the next wave arrives.

There’s even a running joke: In the AI era, if you learn slowly enough, you’ll never have to learn at all.

AI has evolved from a conversational assistant into an executor—capable of browsing the web, reading files, operating desktop software, invoking APIs, and completing multi-step tasks end-to-end.

Microsoft’s 2026 Work Trend Report shows that active Agents within Microsoft 365 have grown 15× year-on-year. Analysis of over 100,000 Copilot conversations reveals that 49% now support cognitive work: analysis, judgment, problem-solving, and creative thinking.

This shift is unmistakable—and likely irreversible. In just a few years, Agent-driven productivity may become a dominant force.

We can distill this transformation into four concise statements:

• AI’s output unit is shifting from an answer to a task.
• AI’s capability carrier is shifting from a prompt to knowledge bases, workflows, and Skills.
• The organization’s fundamental unit is shifting from one person to one person plus a set of Agents.
• Information distribution is shifting from web-page ranking to AI retrieval, citation, and response.

Technology races forward—but throughout our practice, conversations, and writing of AI Leadership, certain questions have remained stubbornly unchanged:

• Does AI truly amplify human productivity?
• Why do so many people still resist using AI?
• What should the human–AI relationship be?
• Is mindset shift alone sufficient?
• Does AI enable intellectual equity—or deepen capability divides?
• As machines grow stronger, what uniquely human value remains?

You’ll find extensive reflections and answers to these in the book—and a 5,000-word summary here: mp.weixin.qq.com

The Essence of Communication

At its core, communication is about building connection—connections that align resources, capabilities, and action.

An individual boosts productivity in three main ways: working alone, leveraging AI, or collaborating with others. The first two improve personal efficiency; collaboration depends entirely on communication.

We can understand communication along two axes:
• Interaction mode: direct (e.g., face-to-face, phone, DM) vs. indirect (e.g., articles, posts, books)
• Reach scale: one-to-one, one-to-many, or small-group multi-to-multi

A blog post or live workshop is one-to-many indirect communication—but it carries high reach efficiency. One piece of content can introduce you to hundreds or thousands simultaneously, establishing a weak tie: no prior conversation, yet sustained exposure builds familiarity with your thinking, judgment, and working style.

When that exposure is long-term, authentic, and valuable, familiarity gradually becomes trust.

That trust shortens the ramp-up time when real collaboration begins—turning weak ties into deep ones.

One recent client followed exactly this path:
A friend followed me on X, recommended my work, and the client then found my WeChat account, read several articles, watched livestreams and public lectures, formed a clear impression—and reached out proactively to propose collaboration. Trust built before the first meeting accelerated everything that followed.

Another example: I recently visited the founder of an outstanding company. We’d met seven years earlier—then had zero contact since. Yet when I asked to meet during a business trip, he accepted immediately. He told me he rarely accepts such requests—but had been following my WeChat Moments and public sharing for years. He felt our values aligned, our questions overlapped, and some of my posts had even sparked new ideas for him.

Consistent, public sharing quietly reshapes weak ties—even those you never directly engage with. It tells people who you are, what you do well, and how you think—laying groundwork for future partnerships before they’re even imagined.

System Refactoring—Powered by AI

Over a weekend, my tech lead and I undertook a full-stack refactor of an aging product—one that now serves several million users through organic growth.

Its underlying architecture and UI hadn’t kept pace. So we rebuilt both the foundation and interface—end-to-end.

In two days, we consumed over 2 billion tokens.

The engineering scope was massive. Under traditional methods, delivering equivalent results would have taken at least three engineers and over a month.

With AI assistance, we completed most of the hardest parts in 48 hours—and concurrently upgraded payment and SMS integration for our conference platform.

It was deeply meaningful work.

Our refactoring logic followed six phases:

1. UI Prototyping & Frontend Engineering
   Static visual mockup → Standalone Vue 3 + Nuxt interactive prototype → Migrated into Laravel main repo as Vue 3 + Inertia + Vite pages

2. Backend Foundation Upgrade
   Laravel 7 → 13; Nginx 1.18 → 1.31.3; plus container, dependency, security, callback, storage, and deployment compatibility fixes

3. Convergence of Dual Tracks
   Parallel frontend/backend design, migration, and testing—with final frontend integration anchored on the upgraded Laravel 13 base

4. Page-by-Page Business Integration
   Each page implemented: real-data fetching, CRUD operations, and feature parity

5. Stage-Gated Testing
   PHPUnit, Vitest, type checking, ESLint, Playwright interaction tests, 6-viewport visual regression, Docker regression, legacy entry compatibility, blind testing, and BT03

6. Unified Deployable Artifact
   Final output: Laravel backend + Inertia bridge layer + Vue 3/Vite public/build assets + legacy-entry compatibility layer

Boundaries and Risk Awareness

I spoke with a friend whose business collapsed rapidly due to regulatory shifts years ago. Since then, he’s been paying down debt.

I asked: “How much is left?”

I assumed maybe a few million.

His answer: nearly 20 million RMB.

I was stunned. It’s incredibly hard.

An entrepreneur’s life may require years—or even a decade—to recover from a single flawed decision.

At dinner recently with several executives from traditional manufacturing and consumer goods firms, one CTO shared his founder’s story: after IPO, abundant cash led to aggressive investments and diversification—seemingly promising moves at the time. But they ultimately saddled both company and founder with crushing debt. Even after selling major assets, the gap remains vast.

I recalled an early mentor—a seasoned entrepreneur—who once told me his founding principle:

First: Running a company means managing risk.
Second: Never forget the first.

Growth determines how fast you go.
Risk control determines how far you go.

Many opportunities warrant risk—but only when you know your boundaries, and consciously balance risk against reward. That dynamic equilibrium may be the single most vital capability for long-term survival.

Yet cultivating boundary awareness—and the discipline to act on it—is harder than most assume.

The Power of AI Search

I was genuinely surprised by this effect from Doubao.

At dinner with nine executives from large traditional manufacturing and FMCG companies (revenues ranging from hundreds of millions to billions), one shared striking data:
Doubao-sourced leads now account for 50% of all inbound leads.

Digging deeper, they realized Doubao was heavily citing their TikTok videos—and they’d been investing steadily in TikTok-based “seed planting” (short-video marketing) for years.

I asked: “What was TikTok’s share before?”
He replied: ~10%.

Doubao’s influence is real—and growing fast.

The Real Value of GEO

What does GEO actually deliver?

1. Short term: Boosts brand awareness, credibility, precise lead acquisition, and downstream conversion rates.
2. Medium term: Stabilizes marketing ROI and brand referrals—and strengthens resilience against algorithmic volatility across AI platforms.
3. Long term: Two underappreciated benefits emerge—and they’re far more consequential:
   • Consistent, accurate, science-backed GEO investment increases the likelihood that your enterprise knowledge gets absorbed into LLM training corpora.
   • In an Agent-dominated future, GEO doesn’t just affect human AI search—it shapes how Agents discover, cite, and act on your information. That volume will dwarf human-scale usage.

Meta-Prompt Upgrade

I’ve fully refreshed my meta-prompt system—designed to help users turn vague ideas into production-ready prompts.

Two core upgrades:
• Quality reinforcement of existing prompts
• High-fidelity generation of first-draft prompts

Underlying principles include: intent recognition, multidimensional prompt diagnosis, silent benchmark learning, domain-specific engineering optimization, quality scoring, static validation, and model-level testing.

Design principles:
• Simple tasks → minimal sufficient structure
• Complex tasks → add tool rules, exception handling, and quality thresholds
• Clearly separate facts, assumptions, and method-level references—no jargon stacking without justification

Usage examples:

1. Paste a rough idea, half-written prompt, or use case—and the system reconstructs the core problem, identifies audience, context, deliverables, constraints, and success criteria—then outputs a ready-to-use prompt.
2. Feed an existing prompt: it preserves core intent and non-negotiable elements, then resolves redundancy, contradictions, ambiguity, and missing components.
3. Paste the entire meta-prompt into any LLM—it guides you interactively with a lightweight template (“Just say one sentence about what you want to do…”).

This system shines in two scenarios:
• You have a fuzzy idea and need it shaped into something executable
• You already have a prompt but want it clarified, hardened, and made more reliable

NewMax AI’s founder, Yangyi, has deployed this meta-prompt system in NewMax—try it yourself.

New meta-prompt resources:

1. Updated GitHub repository:
   yao-open-prompts

2. Full rationale and iteration report:
   doc.laoyao.cn

3. Free NewMax AI download:
   newmax.cc

4. Methodological foundations:

   • OpenAI Prompt Engineering: modular instructions, examples, context, tools
   • OpenAI Prompt Optimizer: data, scorers, optimization loops, retesting
   • OpenAI Prompting: versioning, testing, rollback
   • Anthropic Best Practices: clear instructions, XML separation, long-context structuring
   • Google Gemini Strategies: constraints, formatting, task decomposition, iteration
   • OpenAI Model Spec: hidden reasoning boundaries, instruction hierarchy, low-privilege data handling
   • AI Leadership (our book)

AI Product Launch

I attended NanoWork’s launch event from the perspective of a media practitioner.

Afterward, NanoWork listed my 18 GEO Skills in their “Expert Plaza — Co-branded Experts” channel. I tested it: solid performance.

How to use it:

1. Download NanoWork for desktop: work.n.cn
2. Go to “Expert Plaza → Co-branded Experts”, find “Yao Jingang GEO Optimization”, click “Hire”
3. In the expert chat window, describe your task

Since the Skill pack covers many use cases, you can start by asking: “What GEO capabilities and scenarios does this expert support?”

It currently handles 18 common GEO applications.

Example: I submitted:
“Analyze this website: work.n.cn”

Result: a highly professional diagnostic report (see image). The full report was auto-generated and uploaded here: doc.laoyao.cn

The future of AI adoption isn’t about isolated prompts—it’s about building AI capability systems tailored to your business context. With curated datasets and purpose-built Skills, GEO analysis, methodology design, and principle-based deconstruction become more accurate, objective, and scientific.

On the launch itself—I found Lao Zhou’s talk refreshingly grounded. No grand AI narratives. Just honest, practical answers to real questions:

• Why do enterprises struggle with AI?
• What makes an AI tool actually useful?
• How do products evolve within real business workflows?

1. Current Enterprise AI Adoption Gaps

Lao Zhou frames employee AI usage in four tiers:

• Level 1: Treating AI as a search engine
• Level 2: Using AI as an advisor
• Level 3: Leveraging AI as a production tool
• Level 4: Embedding AI as digital staff—automating entire workflows

He estimates 90% of employees remain at Level 1—typing queries into a search-like box.

Mindset shift matters—but so does tool quality. And true adoption starts at the top: the CEO must adopt NanoWork first. Only then does the organization become a “super-organization.” This mirrors AI Leadership’s central thesis.

2. NanoWork’s Self-Definition

Their tagline says it plainly: “One person, an army. One hundred tasks, entrusted.”
They aim to: help you plan, execute, monitor, develop, earn, and grow.
In enterprise terms: help you find customers, analyze operations, track growth, and cut outsourcing costs.

3. Product Evolution Through Real Work

Some highlights resonated deeply:

• In 5 months, they collected 56,000 real-world user problems—over 11,000 per month
• To validate impact, they deployed 100,000+ Agents into live operations: 150 days, 630 roles, 350 trillion tokens consumed

This reflects a humble, empirical product philosophy: don’t just listen—embed your tool in real work, and measure whether it solves problems, delivers outcomes, and creates value.

From this, Lao Zhou distilled four enterprise AI adoption barriers:
• Cost
• Stability
• Security
• Control

These four words capture the journey from wanting AI to trusting it, using it, and relying on it long-term.

He introduced one key concept: FDE (Follow–Do–Evaluate).

A concrete example: turning meetings into execution loops.
Logic: After the meeting ends, AI doesn’t stop—it follows up, does the work, and evaluates progress.
The full loop:
Recording → Auto-generated minutes → Task extraction → Sync to project system → Owner notified → AI checks progress in 3 days → Auto-alerts on delays → Weekly execution summary

Outcome: AI tracks who owns what, status, blockers, and delays—and surfaces patterns automatically.

The core enabler? AI’s ability to autonomously follow up, remind, and synthesize—powered by NanoWork’s cloud infrastructure.

That’s FDE’s real power: transforming AI from a tool into an organizational execution system.

It sounds simple. But when operationalized, it lifts team efficiency dramatically.