2025: 730,000 Lines of Code, One Clarity

An AI once wrote my obituary. It said: "He mass-produced imperfect but genuine prototypes at a speed most teams couldn't match." It wasn't wrong. But it missed the point. This year wasn't about speed. It was about learning to see clearly — 念头通达.

Prologue — The Obituary

Somewhere in my drafts folder, there's a Twitter thread written by an AI — my obituary. It counted 38 projects, 730,000 lines of code in 13 months, and concluded that "AI was his only colleague." Reading it felt like looking into a mirror held by a machine that understood my patterns better than I did.

I share this not for drama, but because it captures something true about 2025: this was a year lived at terminal velocity, alone, with AI as both tool and companion. What follows is my attempt to make sense of it.

Chapter 1: The Leap — Leaving DJI

I started 2025 as a 26-year-old who had just left DJI.

At DJI, I was an embedded software engineer working on autonomous driving systems and AUTOSAR architecture. Good salary, clear career path, prestigious company. The kind of job your parents brag about at dinner parties.

But I kept thinking about something Clayton Christensen wrote in The Innovator's Dilemma: large companies fail not because they do things wrong, but because they do the right things — for the wrong era. DJI was excellent at what it did. I just didn't want to be excellent at something that felt like it was about to be disrupted.

So I left. Moved between Hong Kong and Singapore. Called myself a "散修" — an independent practitioner, walking an unconventional path. My parents were... concerned.

At this point, I want to be honest: I was basically a beginner at everything except embedded C. I couldn't build a website. I couldn't design a product page. I didn't know React from Vue. I had strong opinions about AI products but zero ability to ship one.

That was about to change.

Chapter 2: Nexting AI — The First Bet

My first real project was Nexting AI — an AI-powered platform for Generative Engine Optimization (GEO).

The thesis was simple: traditional SEO is dying. People don't Google anymore — they ask ChatGPT, Perplexity, Claude. If your content doesn't show up in AI-generated answers, you're invisible. Nexting AI would deploy a "living" Agent on your behalf: monitoring trends, adapting content, driving organic traffic. Set → Forget → Rank!

I designed an ambitious architecture: five AI Agents working in concert — Auto-Update, AI Search Ranking, Content Analytics, One-Click Deploy, and Marketing Advocate. The tech stack evolved through Next.js 16, React 19, LangGraph, WebContainers, and Supabase.

I talked to KOLs in the 出海 (international expansion) community — developers like Idoubi, Randy, James — who gave me valuable feedback. The product was real. The vision was compelling.

But the market didn't care. Not yet.

Chapter 3: The Dark Period

November 2024. I failed to close funding.

I watched competitors raise $50M for ideas that overlapped with mine. My server cost $48 a month. The gap wasn't in technology — it was in narrative, timing, and the uncomfortable truth that a solo founder with no track record is a hard sell.

I could have gone back to a job. The embedded systems market was hot. My DJI background would have opened doors.

Instead, I sat with the discomfort.

During this period, I picked up Kenneth Stanley's Why Greatness Cannot Be Planned. It hit me like a truck. Stanley, an AI researcher at OpenAI, argues that the most important discoveries in history came not from goal-directed planning, but from open-ended exploration. The compass, the printing press, the internet — none were designed to achieve their ultimate impact.

His core insight: pursuing a vision (direction) is fundamentally different from pursuing a goal (destination). Goals constrain. Vision liberates.

I stopped asking "how do I make Nexting AI succeed?" and started asking "what's the most interesting problem I can explore right now?"

That question changed everything.

Chapter 4: Perfect-Web-Clone — The Open Source Breakthrough

The first stepping stone appeared in an unexpected place: website cloning.

I noticed that screenshot-to-code tools (72K+ GitHub stars) were fundamentally fragile. They looked at a picture of a website and guessed the code. That's like trying to understand a building by looking at a photograph — you miss the plumbing, the wiring, the foundation.

I built Perfect-Web-Clone differently: extract the actual source code, then use a multi-agent system to orchestrate the replication. The key insight was that website data is massive — far exceeding the ~200K token context window of any single AI agent. You need multiple agents working in parallel, each handling a piece of the puzzle.

The architecture was elegant:

Web Crawling — comprehensive data extraction
Multi-Agent Replication — coordinated agents handling the complete workflow

Built on Claude Agent SDK with 40+ specialized tools, sandbox-based execution, and structured pipelines for consistent output.

It hit 319+ GitHub Stars. More importantly, it became infrastructure — a foundation for anyone building products like Base44 or v0. I open-sourced everything under MIT.

"好产品的唯一判断标准：产生Magic Moment!" — The only criterion for a good product is creating a magic moment. When someone cloned a complex website in minutes and saw pixel-perfect output, that was the magic.

Chapter 5: The Cat Purr That Started Everything

Here's where the story gets weird.

In the first half of 2025, I built a cat-shaped AI companion toy. A DEMO with a plush body and an AI brain. Its signature feature? Purring. Not random noise — algorithmically generated purrs that changed with the cat's "mood."

I spent weeks studying a purring algorithm created by an Indian developer over a decade ago. I analyzed the acoustics, the frequency patterns, the way real cats modulate their purrs based on contentment vs. attention-seeking.

This led me down a rabbit hole into AI audio. I explored building an AI vocal teacher — a product that could listen to someone sing and give personalized feedback. I tested CNN+RL approaches, generative AI voice comparison, and LLM-based solutions using Qwen2-Audio.

The key insight from all this exploration: audio generation and audio understanding are completely opposite domains. One is generative AI, the other is discriminative AI. Most people conflate them. The real product opportunity was in understanding, not generation.

But the deeper insight was something else entirely. Through building the cat toy, the vocal teacher, and dozens of prototypes, I kept hitting the same wall: AI's bottleneck isn't intelligence. It's access. The AI could analyze, generate, and reason. But it couldn't open your email. It couldn't check your calendar. It couldn't access your files.

AI didn't need to get smarter. It needed permission.

Chapter 6: Pinclaw — 8 Grams to Change Everything

Every design decision in Pinclaw is a deliberate removal.

No screen — you already have a phone and a computer. No camera — that gives you all-day battery and zero privacy concerns. No 5G — your iPhone provides connectivity. No on-device AI — it runs on your own computer through OpenClaw. No proprietary OS — we use Zephyr RTOS, the same system trusted by industrial IoT, medical devices, and automotive.

Saint-Exupéry wrote: "Perfection is achieved not when there is nothing more to add, but when there is nothing left to take away." Pinclaw is that philosophy made physical.

完美不是无可增添，而是无可删减。

The result: an 8-gram clip you attach to your collar, pocket, or bag strap. You talk to it. It talks back. Behind the scenes, your words travel via BLE to your iPhone, then via WebSocket to OpenClaw running on your Mac — where a full AI Agent has real permissions to access your calendar, email, files, and apps.

The specs tell the story of intentional minimalism:

28×20×9mm, ~8 grams
BLE 5.4, Opus 16kbps audio
18-hour continuous usage
Seeed Studio XIAO nRF52840 (64MHz ARM Cortex-M4F)
Zephyr RTOS

My DJI background in embedded systems paid off here. I optimized power consumption from 80mA down to 15mA through architectural choices — not by adding a bigger battery, but by being smarter about every milliamp.

The business model has two paths:

My OpenClaw (Free) — Run your own AI Agent on your Mac/PC. MIT licensed. No monthly fee. No lock-in.
Pinclaw Agent ($29/month) — Pre-configured cloud Agent with GPT-4o, Claude, Gemini. Zero setup.

Hardware pre-order: $99 (down from $129). Both paths use identical hardware. The only difference is where the AI runs.

I learned from Humane AI Pin's catastrophic failure — $699, overheating, bad battery, clunky interaction, stopped selling in February 2025. Pinclaw costs one-seventh the price and does more, because it doesn't try to be a standalone device. It's a voice interface to something much more powerful.

From concept to App Store: 20 days.

Chapter 7: OpenClaw — The Real Innovation

Pinclaw gets the attention, but OpenClaw is the revolution.

OpenClaw is a truly open-source AI Agent platform that runs on your personal computer with full OS-level access. This isn't a chatbot. This is an agent that can:

Open Chrome, log into Gmail, read your emails
Use VS Code to help you write code
Operate Figma, Notion, Slack, Excel — any desktop application
Access your entire filesystem via MCP (Model Context Protocol)
Run Python scripts, manage files, execute complex workflows

The OpenClaw ecosystem has grown to 32,000+ GitHub Stars, supports 20+ message platform integrations, and has an active developer community building "Claw" apps — KiloClaw, clawi.ai, chowder.dev, EasyClaw, ClawSimple, ClawApp.

This architecture solves the fundamental problem I identified through the cat toy and vocal teacher experiments: AI needs real permissions, not more intelligence. Every other AI wearable tries to be a walled garden. OpenClaw is an open field.

Users can choose Pinclaw's managed service or build their own AI backend entirely. Hardware is just the voice entry point. The agent is what matters.

Chapter 8: Writing as Thinking — 18 Posts, 2 Podcast Episodes, and a WeChat Channel

I published 18 blog posts in 2025. Not for clout — for clarity.

The topics tell the story of my year's intellectual journey:

Technical deep-dives: Transformer multi-head attention mechanisms, LLM fine-tuning processes, LangGraph persistence and memory modules, Manus multi-agent architecture analysis, AG-UI protocol guide.

Product thinking: What makes an AI product manager? (Answer: understanding technology trends, not just efficiency gains.) AI travel product analysis. Enterprise SaaS in the age of AI. Cursor product teardown (1/100 in a series I plan to continue).

Philosophy: Why Greatness Cannot Be Planned — my most personal post. The cat purr algorithm — my most unexpected. How to build a singing model — my most technically ambitious.

I also launched "Eric的AI播客" on Apple Podcasts:

EP01: Why "context engineering" is AI's real bottleneck
EP02: Kevin Kelly on the mirror world and the death of privacy

And I ran a WeChat channel called "下一站-Agent" (Next Stop: Agent), publishing long-form technical content for Chinese-speaking developers.

Writing isn't a side project. It's how I think. Every blog post forced me to understand something deeply enough to explain it simply. Every podcast episode made me articulate ideas I'd only felt intuitively. The 18 posts aren't content — they're the intellectual scaffolding that made everything else possible.

Chapter 9: AI Made Me a One-Man Army

This is the chapter that matters most.

In January 2025, I couldn't build a website. I was an embedded C programmer who knew theory about AI products but couldn't ship a landing page. React was foreign. CSS was a mystery. Design was something other people did.

By December 2025, I was shipping multiple websites in a single day.

The transformation wasn't gradual. It was a phase change, and the catalyst was learning to use AI — specifically Claude Code — not as an assistant, but as an extension of my mind.

Here's how my relationship with AI tools evolved:

Phase 1: The Timid Questioner (January–March) I'd ask Claude Code one question at a time, wait for an answer, read it carefully, then ask the next question. I treated it like a search engine that could write code. Slow. Cautious. Inefficient.

Phase 2: The Delegator (April–June) I started giving Claude Code larger tasks. "Build me a blog component." "Set up the API route." I learned to provide context, review output, and iterate. Faster, but still sequential.

Phase 3: The Orchestrator (July–September) This is where everything changed. I stopped thinking of AI as a tool and started thinking of it as a team. I'd spin up multiple agents in parallel — one working on the frontend, one on the backend, one researching a technical question. I became the director, not the actor. The product manager, not the engineer.

Phase 4: The One-Man Army (October–December) By year's end, I was operating like a small studio. In a single day, I could:

Ship a complete website from design to deployment
Create 3D models for hardware prototypes
Generate and edit AI videos for product demos
Record, edit, and publish podcast episodes
Write and distribute content across multiple platforms
Debug firmware, optimize BLE connections, and test hardware

Skills I acquired through AI collaboration in 2025:

Web development: React, Next.js, TypeScript, TailwindCSS — from zero to production
3D modeling: Designing hardware enclosures, printing prototypes on my desk
Video production: AI-generated product videos, editing, post-production
Audio engineering: Podcast production, audio processing algorithms
Design: UI/UX design, product pages, marketing materials
DevOps: Cloudflare Workers, Docker, CI/CD pipelines, R2 storage
Hardware: BLE firmware, power optimization, Zephyr RTOS

The key insight isn't that AI tools are powerful. It's that AI changes what "skill" means. I don't need to master React — I need to master directing an AI that knows React. I don't need to learn Blender — I need to articulate what I want in 3D space. The bottleneck shifted from "can I do this?" to "can I describe what I want clearly enough?"

This is what I mean by using Claude Code like a product director. I don't write most of the code. I architect the system, define the requirements, review the output, and make judgment calls. I run parallel workstreams the way a PM runs parallel sprints. The AI handles execution. I handle vision and quality.

730,000 lines of code in 13 months. Not typed — directed.

Chapter 10: The Infrastructure of One

Being a one-man army requires infrastructure. Here's what I built:

Automated content distribution: Write once, publish everywhere. A single blog post gets adapted to Twitter threads, 小红书 cards, WeChat articles, and Zhihu answers — each version optimized for the platform's native audience.

Twitter automation: Connected to the API (Free Tier, 1,500 posts/month). Claude Code writes and schedules posts based on my content calendar.

Podcast pipeline: Python scripts for audio processing, automated RSS feed generation, Cloudflare R2 for hosting, Apple Podcasts distribution. End-to-end: write a script → generate audio → publish → distribute.

Multi-platform deployment: Cloudflare Workers, Vercel, GitHub Pages. I can go from idea to live URL in under an hour.

Hardware prototyping: 3D printer on my desk for rapid iteration. Arduino and BLE experiments for firmware testing. The Pinclaw hardware went through dozens of physical prototypes.

All of this runs on roughly $48/month in server costs. The most expensive tool in my stack is my time.

Chapter 11: What I Got Wrong

I have 6 Twitter followers.

Let that sink in. 730,000 lines of code, 319 GitHub stars, an App Store listing, 18 published blog posts, a podcast, multiple products — and 6 people follow me on Twitter.

I told myself it was intentional. "Building, not broadcasting." That was a lie I told myself to avoid the discomfort of self-promotion.

Here's what else I got wrong:

50 Douyin video ideas, zero recorded. I have a detailed list of topics — "The biggest lies about solo entrepreneurship," "How to pay yourself your first salary" — sitting in a folder, gathering digital dust. Knowing what to say and actually saying it are different skills.

小红书 content, never published. Same story. Drafted, planned, never shipped. The irony of someone who preaches "ship fast" having an entire unpublished content library is not lost on me.

Pinclaw's honesty problem. Every product image on the website is AI-rendered. "AI-generated for illustration purposes. Actual product may differ." The hardware hasn't shipped yet. I'm in pre-order stage. This is either bold vision or premature marketing, and I won't know which until units are in people's hands.

Solo by choice or by limitation? "AI is his only colleague" sounds romantic in an obituary. In reality, there were nights at 00:53 AM fixing relay reconnect timeouts where a co-founder would have made all the difference. I optimized for independence when I should have optimized for leverage.

The funding gap. Competitors raised $50M. I had $48/month in server costs. I framed this as a virtue — lean, bootstrapped, efficient. But sometimes you need capital to move fast, and my refusal to play the funding game may have cost me a year of runway.

Chapter 12: The Books That Shaped My Thinking

Four books rewired my brain in 2025:

"Why Greatness Cannot Be Planned" — Kenneth Stanley The most important book I read this year. Exploration over exploitation. Stepping stones over milestones. The cat purr algorithm led to audio AI led to the realization that AI needs permissions led to Pinclaw. None of this was planned. All of it was connected.

"The Innovator's Dilemma" — Clayton Christensen Why I left DJI. Why I don't fear competing with funded startups. Incumbents optimize for their existing value network. Disruptors find new ones. I'd rather find a new network than fight for position in an old one.

"Hackers and Painters" — Paul Graham The creator's mindset. Programming as craft, not industry. The importance of making things, even imperfect things, over theorizing about perfect things.

Saint-Exupéry (via design philosophy) "Perfection is achieved not when there is nothing more to add, but when there is nothing left to take away." This sentence is carved into every Pinclaw design decision.

Chapter 13: Life — Freedom & Love

The cover image of this post shows me swimming. It's not accidental.

2025 was lived between Hong Kong and Singapore — two cities that feel like they're racing toward different futures. Hong Kong is dense, electric, full of friction. Singapore is optimized, smooth, almost too efficient. I need both. The friction generates ideas. The efficiency lets me ship them.

My personal pursuit is "念头通达" — clarity of mind. Not enlightenment in some spiritual sense, but the ability to see a problem clearly, strip away the noise, and act on what matters. Every project, every blog post, every late-night debugging session was, in some way, practice for this.

The swimming metaphor works: you can't think your way through water. You have to feel it. Move with it. Let it hold you up while you do the work.

I'm 26. I have no funding, no team, no safety net. What I have is an 8-gram device, a clear vision, and the hard-won ability to build anything I can imagine.

That's enough.

Epilogue — 2026: No Plans, Just Stepping Stones

Pinclaw launched 3 days ago. Zero press coverage. No viral moment. No TechCrunch headline. Just a quiet listing on the App Store and a GitHub repository that anyone can fork.

I don't know if it'll work. I don't know if 2026 will be the year Pinclaw finds its audience, or the year I learn another hard lesson about hardware manufacturing, market timing, or the gap between vision and reality.

What I do know is this: in January 2025, I couldn't build a website. In December, I shipped hardware, software, podcasts, open-source tools, and 18 blog posts that I'm genuinely proud of. AI didn't just help me build faster — it transformed what I believed was possible for one person to accomplish.

Kenneth Stanley would say: don't set goals for 2026. Set a direction. Follow the interesting stepping stones. Trust that the cat purr algorithm might lead to the next big thing, because it already did once.

So here's my non-plan for 2026:

Keep building. Keep subtracting. Keep exploring.

And maybe, finally, record those 50 Douyin videos.

越挫越勇 — Fail forward, always.

I can't hold back the earth-shattering ideas in my head, so I choose to keep making mistakes and bring them to life, one by one.