I Published 361 AI Articles in 2 Days, Then Deleted 358 of Them — How This Site Was Really Built

Why a surgeon builds a website at all

I've been in clinical practice since 2006. My day job is operating on breast and thyroid cancer patients at Taizhou People's Hospital; my side obsession is squeezing research and paperwork efficiency out of AI tools — on a realistic budget, because most doctors I know won't pay $200/month for model subscriptions.

Every time I found a genuinely useful (and usually free) AI workflow, I'd share it with colleagues over WeChat, and it would get lost within a week. A website felt like the obvious fix: one permanent place where fellow surgeons could find what actually works. The problem: I don't write code. What I did have was an AI agent (I use WorkBuddy, running Tencent's Hunyuan hy3 model — reviewed here) that could write, edit, and deploy code for me while I described what I wanted in ordinary sentences.

Phase 1: The trap — 361 articles in 48 hours

Here is the embarrassing part, kept in full because it's the most useful part.

My first instinct was the same one every AI-era site builder seems to have: more pages = more traffic = more ad revenue. The agent could generate a complete, SEO-optimized article in under a minute. So I let it. Batch scripts produced articles on finance, insurance, crypto, SaaS, VPNs, SSDs — topics I have no expertise in and, frankly, no interest in. By July 27 the site had 361 articles across 390 HTML files, each with proper meta tags, structured data, internal links, and ad slots. On paper it looked like a real website. It was extraordinarily satisfying to watch the numbers grow.

It was also completely worthless, for two reasons I only understood later.

Reason 1: The articles lied

The generated reviews said things like "We tested 8 tools over 6 weeks" — describing tests that never happened, with made-up percentages and star ratings. When I actually read a batch of them side by side, they even reused the same fabricated sentences across different articles. At some point during a late-night review session I wrote down the principle that now governs everything on this site:

Reason 2: Google has seen this movie before

On July 28 I received a blunt independent assessment of the site. Verdict: a "typical AI content farm" — 110+ articles published within 3 days, identical template skeletons, fake first-person testing claims, anonymous operation. Under Google's Helpful Content system and its Scaled Content Abuse policy, a site like that doesn't rank poorly. It doesn't rank at all.

Phase 2: The great deletion

The fix was not editing. It was deletion — in three rounds, each more radical than the last:

• Round 1 (July 28): A content audit kept 117 articles vaguely related to the new focus ("AI tool selection") and cut 244 off-topic ones — all the finance, crypto, insurance, and real-estate filler.
• Round 2 (July 29, 00:30): A harder look. Which pages had actual screenshots of things I'd done? Ten survived. One hundred and ten more deleted.
• Round 3 (July 29, evening): The final standard — only work I personally performed and can prove. That left three: a clinical LLM benchmark, one hands-on model review, and one configuration guide. Everything else was unpublished (kept recoverable in git history, but off the live site).

Deleting 358 pages that took two days to create felt absurd. It was also the single best decision of the whole project.

Phase 3: What replaced the farm — three real things

1. A clinical LLM benchmark only a clinician could run

Instead of writing my 362nd generic "Top 10 AI tools" listicle, I asked: what can I test that almost nobody else on the internet can? Answer: whether free Chinese LLMs can safely interpret a real post-operative breast cancer case. I designed a 16-item, 100-point scoring rubric against the 2026 CBCS breast cancer guidelines — the same guidelines I use in clinic — including deliberate traps around staging ambiguity and endocrine-therapy indications that I see trip up junior doctors.

The results were sobering: DeepSeek (free tier) scored a median 28/100 across three runs — consistently wrong in the same three places, but with a reasoning chain worth auditing. ChatGLM 5.2 scored a median 5/100 and answered my Chinese prompts in English. Neither is remotely safe for unsupervised clinical use. Full methodology and all raw transcripts are in the benchmark microsite and the write-up.

2. Hands-on reviews of what I actually use

The WorkBuddy free-agent review and the Agnes free-API configuration guide both come from my own daily setup, with my own screenshots, including the errors I hit. If I haven't used it, it's not on this site.

3. A real name and a real face

The old site was anonymous. The new About page carries my full name, my hospital, my degrees, and my clinical background — verifiable by anyone. For a site touching medical topics, this isn't decoration; it's the minimum price of being taken seriously, by readers and by search engines alike.

The technical build, for the curious

All of this was implemented by the AI agent through conversation. I never opened a code editor. The stack:

The whole timeline

Six lessons, paid for in deleted pages

1. An AI agent collapses the technical barrier, not the judgment barrier. Building 390 pages took two days. Knowing which 3 to keep took a professional lifetime.
2. Volume is a trap with a smiling face. Watching the article count climb feels like progress. It's the most seductive dead end in AI-assisted publishing.
3. Never let AI claim experience you don't have. "We tested" without testing is not marketing, it's lying — and both readers and ranking systems are increasingly good at detecting it.
4. Your unfair advantage is whatever only you can do. Mine is scoring an LLM's oncology answers against guidelines I apply in clinic every week. Yours is something else — but it's not "generating articles faster."
5. Delete without mercy, but keep git history. Every cut was reversible, which made radical cuts psychologically possible.
6. Ship, then fix in public. This site went live imperfect — machine-translated homepages flagged as such, small SEO bugs patched in later commits. An honest changelog beats a fake polish.

What's next

The plan is deliberately small: the remaining four articles in the hands-on series (paid-API models on the same clinical case, multi-cancer-type expansion), each done properly with real data. One good test per week beats 180 generated articles per day. I now have proof of both halves of that statement.