330 Naver Blogs Scraped + Full SEO/AEO Overhaul — 3 Sessions, 604 Tool Calls
The longest session today ran 13 hours and 13 minutes. 422 tool calls. I parsed 330 Naver blog posts, designed a dental blog auto-generation pipeline from scratch, and simultaneously overhauled SEO/AEO for jidonglab and spoonai.
TL;DR Session 2 strengthened search visibility with llms-full.txt, sitemap, and structured data. Session 3 analyzed 330 Naver blog structures and produced 8 HTML blog card templates based on S-tier patterns.
The Telegram 409 That Kept Coming Back (15 min, 14 tool calls)
Session 1 lasted 15 minutes. Same problem as yesterday, same day.
I had just installed the fakechat plugin, so Claude initially connected it to the Telegram issue — they were completely unrelated. The actual cause was in the MCP server logs:
409 Conflict: Terminated by other getUpdates request;
make sure that only one bot instance is running
Every time a Claude Code session restarts, the Telegram MCP server spawns a new instance. If the previous process is still alive, Telegram’s Bot API throws a conflict — only one getUpdates long-poll connection is allowed at a time. Releasing the polling lock and restarting via /reload-plugins fixed it.
When Telegram MCP stops responding, check for duplicate processes before touching config.
ps aux | grep "bun server.ts"tells you how many instances are running.
Full SEO/AEO Overhaul (42 min, 168 tool calls)
Session 2 started with fixing broken images on spoonai and ended with a full AI search visibility strategy rollout.
Why the Images Were Broken
Four article thumbnails had been saved as HTML files. anthropic-mythos-01.jpg was 919 bytes — if a .jpg is under 1KB, it’s HTML. The image generation script was silently swallowing Gemini API errors and writing the raw HTTP response body to disk.
Two fixes: removed the image field from frontmatter on 8 broken articles — a missing image is better than a broken one. Added response file size validation to SKILL.md.
Four duplicate articles existed (Harvey AI was collected three times across 03-27, 03-29, 03-30) because deduplication was checking titles, not URLs. Slightly different titles slip through.
llms-full.txt and AI Crawler Access
One prompt — “apply every effective strategy to maximize AI search result exposure” — had Claude modify 14 files and create 7 new ones.
Three core changes:
/llms-full.txt route: a new endpoint that exports full markdown content structured for LLM crawlers. Perplexity, ChatGPT, and other AI search engines prioritize this file when indexing.
AI crawler allow rules in robots.txt:
User-agent: GPTBot
Allow: /
User-agent: Claude-Web
Allow: /
User-agent: PerplexityBot
Allow: /
JSON-LD structured data in Base.astro and PostLayout.astro: Article, BreadcrumbList, and WebSite schemas. sitemap.xml.ts replaces Astro’s auto-generated sitemap — each post’s date now maps to lastmod.
The Dental Blog Pipeline (13 hours, 422 tool calls)
Session 3 was the main event. Read 170, Bash 102, Edit 75, Write 48.
From One URL to 330 Blog Posts
It started with a single Naver blog URL. That turned into “compare top 10 Bundang implant posts,” which became a pipeline collecting the top 3 results across 16 regions × 7 procedures — 336 posts total.
Naver blogs use an iframe structure. The main URL is a 3KB wrapper; the actual content lives at the PostView URL. Claude’s first Bash call pulled 292KB of HTML.
macOS grep doesn’t support -P (Perl regex), so Claude switched to Python immediately after seeing the error — no instruction needed:
import re
from collections import Counter
with open('sample_postview.html') as f:
html = f.read()
classes = re.findall(r'class="([^"]*)"', html)
counter = Counter()
for c in classes:
for cls in c.split():
counter[cls] += 1
print(counter.most_common(20))
se-text-paragraph appeared 272 times, se-module-image 84 times. The Naver SmartEditor 3 structure pattern was now visible.
Working Around contextzip
Naver API responses were disappearing. contextzip was compressing them to save tokens. The workaround: save to file first, then read with the Read tool.
# contextzip compresses responses — results vanish
# save to file and read separately
curl "https://openapi.naver.com/..." > /tmp/naver_results.json
# then use the Read tool
The same pattern handled all 300 HTML parse operations.
Background Task + /compact
Collecting 330 blogs ran as a Background Task — other work continued while collection ran in parallel. That’s why TaskUpdate appears 33 times in the log.
As the session stretched on, context hit its limit. Running /compact preserved the critical context (S-tier patterns, analysis state, next steps) in the summary, and the session continued without losing its thread. One /compact across 13 hours. A good summary means no interruption to flow.
S-Tier Patterns and Image Caching Strategy
The top-ranked Bundang implant blog had 27 components, 36 images, and 26 slides. Analyzing all 330 posts the same way, the top 20% shared patterns that became BLOG-DESIGN-GUIDELINE.md, then a Claude Code skill. Now /write-dental-blog generates guideline-based posts in one command.
The image system was designed around pre-generation rather than on-demand: generate 30–50 base images with Gemini API upfront, cache them in assets/manifest.json, and use cache-first lookup during card generation.
def get_or_generate(prompt_key: str, output_path: Path) -> Path:
if output_path.exists():
return output_path # cache hit
return generate_with_gemini(prompt_key, output_path)
The Quality Feedback Loop
After the first draft of blog cards came back:
font / dark navy tone / logo image isn't working properly / the generated images are all bad
rewrite the logic from scratch.
75 of the Edit calls were concentrated in this feedback loop. The final output: 8 card types for an implant post — title, checklist, procedure steps, info+image cards (bone graft, digital guide, aftercare), doctor profile, legal notice, CTA.
Tool Call Breakdown
Total: 604 tool calls (3 sessions)
- Read: 210 (35%) — HTML parse results, config files
- Bash: 164 (27%) — Python scripts, API calls
- Edit: 101 (17%) — HTML card revisions, SKILL.md updates
- Write: 53 (9%) — new file creation
- TaskUpdate: 33 (5%) — background task status
- TaskCreate: 15 (2%) — background task spawning
- Agent: 6 (1%) — parallel subagent processing
Read 35%, Bash 27%. Exploration (Read + Bash) outpaced implementation (Edit + Write) by 2.4×. Makes sense for a session that was primarily parsing 300 HTML files.
What This Session Confirmed
When contextzip compresses API responses in the environment, saving large results to a file first and reading with the Read tool is the safe pattern. Diagnosing disappearing responses costs time.
Visual quality judgment requires a human. “The images are all bad” can’t be encoded in a script. The 13-hour session stretched because the feedback loop ran multiple times. Next time: share reference images at the draft stage before generating anything.
The Background Task + /compact combination holds up for long sessions. Work in parallel while collection runs, compress when context fills, continue without losing state. That’s how a 13-hour session finishes as a single coherent thread.
More projects and build logs at jidonglab.com
Comments 0