TL;DR: Gave a Claude Code Agent a 500K-line keyword CSV, a Destoon CMS server, and 8 hours. It produced 33 articles (130K Chinese chars), a keyword scoring algorithm, a quality scoring system, and recovered from 3 production outages. Average article quality score: 89.5/100.
The Setup
- Input: 500,001-line CSV from the 5118 keyword tool (GBK-encoded, HVAC thermostat niche)
- Target: www.syuan.vip — a Destoon CMS platform on AliCloud ECS
- Agent: Claude Code Explore Agent (read-only exploration + Python execution + SSH to server)
- Constraint: Zero human intervention during execution
What It Did
Phase 1: Keyword Analysis (30 min)
Built a scoring algorithm grading all 500K keywords across 3 dimensions: domain relevance, search intent, and long-tail features. Discovered that 70% of high-value keywords were tutorial-intent (\"how to wire\"), only 5% commercial (\"price\"). This drove the whole content strategy.
Phase 2: Content Planning (45 min)
Wrote 8 pillar articles first, then built a coverage analyzer (regex semantic domains) to find keyword gaps. Found 10 uncovered domains like electric heating devices (2,789 keywords), sensor selection (1,068), communication protocols (769), etc. This drove 3 more batches to reach 33 articles total.
Phase 3: Writing + Publishing (5 hours)
Built a pipeline: Markdown → HTML converter → POST to CMS API → PHP scripts for TOC generation, CTA injection, internal link conversion, Schema JSON-LD injection, sitemap updates, and IndexNow submission. ~15 seconds per article end-to-end.
Phase 4: Quality Control (1 hour)
Defined a 6-dimension scoring rubric (25+20+20+15+10+10 = 100 points) and built a Python crawler that scores every live article automatically. Average: 89.5/100. 20 out of 33 articles scored A-grade (≥90).
The Biggest Disaster
Tried to inject OG tags into the Destoon template header.htm. Accidentally removed one {/if} from a template with 54 balanced pairs. Entire site returned 0 bytes. Recovered from the 20260115 system backup after 3 restore cycles.
Lesson: Never touch {if} structures in legacy CMS templates. Inject at the PHP logic layer, not the template layer. And always check the compiled template cache before testing.
What Worked Surprisingly Well
- The scoring algorithm was eerily accurate — the \"电地暖温控器接线图\" keyword it ranked #1 is genuinely one of the highest-volume queries
- Schema injection via direct DB writes was bulletproof once we gave up on template-level injection
- The quality scorer caught real issues (articles with 2 internal links instead of 3-4, missing FAQ patterns)
Numbers
| Metric | Value |
|---|---|
| Articles written | 33 (+ 2 blog posts) |
| Total characters | 130K Chinese |
| Average quality score | 89.5/100 |
| Production outages caused | 3 |
| Backup restores | 3 |
| Time to recover | ~45 min total |
Limitations
- No product images (wiring diagrams, product photos need human addition)
- Price data from public knowledge, not actual sales data
- Legacy CMS template adaptation consumed ~25% of total time
Verdict
AI agents are incredibly effective for informational SEO content (how-to guides, knowledge bases, long-tail coverage). Not suitable for content requiring first-party data, genuine user experience, or original photography.
Would love to hear if anyone else has tried similar experiments.
*Site: https://www.syuan.vip/course.html | Full source and scoring code in repo*