Revision note: This post was first published on 2026-07-11 as just an outline, then expanded into its current version on 2026-07-27. The publish date keeps the original date, so you'll see links to articles published after this date.
I'm not a trained programmer. My core skill isn't traditional coding experience — it's quickly assembling business problems, AI tools, and automation pipelines into working systems: breaking messy requirements down into flows, tables, interfaces, and state machines, using AI-assisted development to compress the implementation cycle, and then keeping them stable in production.
The seven cards on the homepage don't have room for the details — this post is the long version.
The starting point · Getting started: my personal Vibe Coding beginnings
I was born after 2005, and I'm a clinical medicine undergrad. Before early 2026, my entire relationship with "programming" was: no relationship at all.
The first thing I built was called Linglan, a medical-knowledge chat webpage: I used DeepSeek plus Volcano Engine's vector knowledge base, fed it PDFs of our textbooks, then used a system prompt to make it handle pre-reading, mind maps, review, and mnemonic tricks. The webpage and RAG were built with AI's help; what I brought was the textbooks and med students' own study methods. Later a guy told me this thing couldn't beat Doubao, and college students weren't willing to pay for it anyway. I was a bit resistant at first, but looking back now, he was right — Linglan was just a demo, with no real users. I go into more detail on this in another article.
But Linglan also confirmed one thing for me: not knowing how to write code doesn't mean you can't start building things.
January (01) · Exploration: falling down the agent rabbit hole
In late January, I typed my first-ever command to Claude Code in a terminal. After that, I was in Claude Code and OpenClaw every day, and from then until now I've racked up thousands of conversation sessions.
The first decent thing that came out of all that soaking was an open-source project in early February: I designed a memory system for agents using a neuroscience-inspired approach — layered storage, facts kept separate from beliefs, memories that decay and also consolidate — and put it on GitHub, where it picked up a few dozen stars. Looking back, the benchmark numbers I gave at the time weren't rigorous enough, and I later spelled that out clearly in the archived write-up. But this project made me realize that medical training's way of thinking is transferable: differential diagnosis, hierarchical classification, handling uncertainty — all of it applies to looking at these systems.
February · Hands-on: into a real business environment
An opportunity brought me embedded on-site at a medical aesthetics chain, doing AI implementation. Three workstreams kicked off almost simultaneously: WeCom AI customer service, Douyin (TikTok's Chinese sibling) lead syncing, and a Xiaohongshu (RedNote) content matrix.
This is where I first learned what "real delivery pressure" actually means: getting a demo to work is not the same as a system that keeps working. It also forced me to be honest about the split: the code was written with AI's help; what I'm responsible for is requirements, workflows, testing, acceptance, and iteration. I won't pretend I can write code, but whether a system ships, whether it actually runs, and how problems get caught — that's on me.
March · Systematizing: from "it runs" to "there's a method"
As the projects piled up, I forced myself to start writing experience down: decision logs, lists of pitfalls, content strategy. The pitfalls I hit later got pinned up on the site's pitfalls wall.
March also brought my first chances to speak on stage — I gave a talk at a local AI-themed event on how to actually implement "digital employees." When I later put the talk transcript up on the site, I went back and tightened up the project status claims and efficiency numbers that the original draft had overstated — what you say to an audience has to hold up to scrutiny too.
April · Delivery: waking up to production discipline
Production discipline is something an incident beat into me. The scheduled job for lead syncing failed silently 22 times, and for three weeks nobody knew — including me. Since then I've made three things standard equipment for every automation project: alerts on failure, heartbeats on success, and data that gets read back to verify. The full retrospective is in this post. These days, the first question I ask when evaluating any system is: when it goes down, how do you find out? If I can't answer that, it doesn't count as delivered yet.
In April I also received a training-instructor appointment letter from the JD.com (Zhengzhou) Merchant Service Center — AI Implementation in Practice track, appointment running through November — a piece of visible outside validation.
May–June (05–06) · Lying low: back to school for finals
Finals season, back to school. Medicine is my fallback — I can't let it slip. AI development basically stopped completely for these two months; it was just classes, exam prep, and exams.
Technically, these two months are a blank. I'm not going to dress that up or hide it: keeping up with my coursework is itself one of the delivery standards I've set for myself.
July · Back at it: full-time again
Summer break, full-time again. What got done this month: this personal site went live, then got "engineered" by me and AI together to the point of having test gates, publishing standards, and content red lines; the customer service system kept iterating in production; and I wrote my first batch of engineering retrospectives — the silent-failure post was written this month.
The goals going forward are specific, with dates set — when the time comes, the results will speak for themselves.
Pitfalls and reflections
The lessons I've knocked into from real projects, and the things I've figured out, are all pinned up on one wall: the full pitfalls wall is here →
I try to make sure every number and claim on this site can hold up to scrutiny. If you're working on something similar, or have a real business you want to talk about, my contact info is at the bottom of the homepage.