I Built a Full SEO Content Pipeline with an AI Agent — 33 Articles, 0 Human Touches

The Experiment

I gave a Claude Code agent 500,000 keywords, a CMS server, and 8 hours. The goal: build a complete published content library with zero human intervention.

Here\'s what happened.

The Tech Stack

• Keywords: 5118 tool CSV — 500K rows, GBK encoding
• CMS: Destoon (legacy PHP platform)
• Agent: Claude Code Explore Agent
• Publishing: Custom Python + PHP pipeline
• Quality: Self-built scoring crawler

Phase 1: Keyword Scoring

Problem: Only 47 out of 500K keywords had search volume data. Can\'t rely on missing metrics.

Solution: Built a scoring algorithm with 3 dimensions:

def score_keyword(kw):
    score = 0
    # Business relevance: 10-35 pts
    for term in [\'地暖\',\'壁挂炉\',\'分水器\']:
        if term in kw: score += 30
    # Intent signals: 5-28 pts
    for term in [\'怎么接线\',\'故障\',\'价格\']:
        if term in kw: score += 25
    # Long-tail features: 3-12 pts
    score += min(len(kw) // 4, 12)
    return score

Result: 70% of top keywords were tutorial-intent. Content strategy: informational, not commercial.

Phase 2: Coverage Analysis

Wrote 8 pillar articles, then built a coverage analyzer using regex semantic domains. Found 10 keyword gaps driving 3 more content batches.

Phase 3: The Pipeline

Markdown → YAML strip → HTML convert →
POST API → TOC generation → CTA injection →
Internal link conversion → Schema JSON-LD →
Sitemap update → IndexNow submission

15 seconds per article. All PHP enhancement scripts run server-side.

Phase 4: The Crash

Tried modifying Destoon\'s header.htm template. Accidentally removed one {/if} from 54 balanced pairs. Entire site → 0 bytes.

Recovered from backup. Lesson: never touch legacy CMS template logic. Inject at the data layer.

Results

• 33 articles, 130K Chinese characters
• Average quality score: 89.5/100
• 20 A-grade (≥90), 8 B-grade, 5 C-grade
• Live at www.syuan.vip/course.html

What I Learned

AI agents excel at: informational content, keyword analysis, pipeline automation, quality scoring.

AI agents struggle with: legacy CMS quirks, first-party data gaps, image generation.

Full source: GitHub