How a non-programmer built a growth flywheel with AI. No ads, no cross-promos, no giveaways.
January 26: 10,000 followers. Early March: 15,000.
Forty days. No paid promotion, no mutual follows, no giveaways.
This isn't a growth playbook. I'm not qualified to write one. It's a retrospective: what I did, what worked, what didn't, and one thing I didn't expect: the thing that actually grew my audience wasn't "making content." It was building systems.
Getting to 10K took most of the previous year.
The daily routine: manually scrolling X, Telegram, and WeChat feeds. Spot good English content, translate it with commentary, occasionally share my experience running quant strategies on Polymarket. Quality was fine, but growth was slow because the bottleneck was obvious. There's a hard limit on how much one person can read and write per day.
I was spending more time hunting for topics than actually writing.
Mid-February, I got fed up with switching between five apps and decided to build my own content aggregator.
I'm not a programmer. Before late 2025, I couldn't write Python. Every line of code was "vibe coded" with Claude Code—I described what I needed, AI wrote the code, I tested and iterated.
The tool was called x-reader: it pulls content from WeChat, Xiaohongshu, Telegram, and X simultaneously, with AI auto-prioritizing relevance. Built purely for personal use, but I casually tweeted the build process.
481 likes, 75 comments.
The replies were full of "can you open-source this?" I spent a day cleaning up the code and shipped it. Posted again: "My first open-source project. Couldn't write Python a few months ago."
561 likes, 110 retweets.
I didn't realize it at the time, but looking back, that was the moment the flywheel started spinning.
After open-sourcing x-reader, I shifted more energy toward building workflows. Not to create content—to improve my own efficiency. But every time I finished building something, sharing the process was naturally a tweet.
The Claude Code memory system is the clearest example. It started as a simple CLAUDE.md config file. After a month, it wasn't enough—AI's context window loses information across sessions. So I added today.md for daily progress, patterns.md for lessons learned, rules/ for behavioral guidelines, docs/ for on-demand documentation.
Without planning it, I'd built a three-tier memory architecture—hot (daily), warm (project-level), cold (vector search)—borrowed from how databases handle caching.
I tweeted each step. Codex cross-review: 901 likes. Agent memory system: 407. Obsidian second brain: 293. Skills ecosystem: 343.
On March 4, I read a paper called "Everything is Context" about AI agent context engineering. It triggered a strange feeling—every concept in the paper was something I was already doing. I just didn't know the names:
My today.md → the paper calls it Scratchpad. My MEMORY.md → Fact Repository. My rules/ → Persistent Directives. My three-tier retrieval → Hierarchical Memory.
Three months of building, and I could never articulate what I was creating. The paper gave it a name.
I posted a side-by-side mapping of my file structure against the paper's concepts. Over 2,000 bookmarks, nearly 2,000 likes—bookmarks exceeding likes means people weren't just appreciating it, they were saving it to replicate.
The paper's author, Robert Mao, commented that it was "the clearest comparison with mainstream framing" he'd seen.
This wasn't "read the paper, then build." It was build for three months, then the paper confirmed the direction. Practice first, theory second—that sequence matters.
The fastest growth didn't come from "the best content"—it came from open-source and deliverables. The x-reader open-source posts totaled 1,000+ likes. The workflow framework release: 628 likes. These posts converted followers at a far higher rate than opinion pieces.
Viral translations are outliers, not part of the flywheel. A February 14 translation about AI-generated TikTok content hit 1,527 likes, but that's not replicable—you need the original author to write something great, and you need to spot it first. That's luck, not system.
What didn't work: Paper teardowns without personal angle (62 likes). Quotes without hands-on experience (36 likes). Posting too many times a day, diluting each other.
People ask how I manage daily posting while maintaining quality. The answer is I'm not "doing content." I'm building tools, assembling systems, running strategies—that's my actual daily work. Sharing is just making the process public.
AI lets a non-programmer build things that used to require a team. x-reader, the memory system, multi-model cross-validation, the automated content pipeline—all built by one person and Claude Code.
Follower count is one number on the dashboard. What's actually valuable is the system itself—it keeps me learning and building every day. Growth is just a side effect.
Next stop 20K. But honestly, the follower number matters less than what gets built on the way there.