Full Content Automation: One Script Generates, Formats, and Publishes in Two Languages

Generating content with AI is table stakes. Generating it in two languages, formatting it for two platforms, and publishing it automatically — that is the system I built. One script execution produces 8 files (4 stories x 2 languages) and pushes them live.

The Pipeline

jidonglab.com is a portfolio and tech blog. It publishes 4 AI news items daily, targeting both English and Korean readers. The old workflow was manual: pick stories, write them, format frontmatter, publish. The new workflow is a single command.

Core requirements:

• 4 AI news articles generated daily
• English and Korean versions simultaneously
• Platform-specific frontmatter and formatting for each language
• Claude API for high-quality content generation

Prompt Engineering Drives Quality

The difference between usable and unusable AI-generated news is entirely in the prompt constraints.

The prompt that works:

"Generate 4 AI news articles for 2026-03-14. Each article:

• Title: SEO-optimized, under 50 characters, include primary keyword
• Body: 300-500 words, technical accuracy first
• Sources: trusted tech media only (TechCrunch, The Verge, Wired)
• Tone: neutral, no exaggeration, fact-driven
• Category: AI/ML, Robotics, or Enterprise AI
• Frontmatter: title, date, category, tags"

The prompt that does not:

"Write some AI news"

The working prompt specifies character counts, structure, source constraints, and output format. The failing prompt specifies nothing and produces inconsistent results every time.

Tone Differentiation Is Not Translation

English and Korean versions need different tone because the audiences are different:

English (jidonglab.com):

"Generate AI news in professional English tone. Target: tech professionals, startup founders. Style: concise, data-driven, business impact focused. Avoid hype and marketing speak. Include specific numbers and technical details."

Korean (DEV.to):

"Generate AI news in Korean for the developer community. Tone: friendly but professional, informal speech. Style: practical perspective, developer-relevant content. Include tech stack details, open-source status, Korean market impact. Avoid translated-sounding Korean."

Same stories, different framing. The English version is 300-400 words and business-focused. The Korean version is 400-600 words with deeper technical context.

Multi-File Generation With Context Coherence

The script generates 8 files per run. The generation order matters for Claude's context management:

# 1. Generate 4 English originals first
generate_english_news() {
  # Request all 4 from Claude in a single call
}

# 2. Localize each English article to Korean
translate_to_korean() {
  for file in $english_files; do
    # Not translation — localized rewriting
  done
}

English first, then Korean localization. This way Claude maintains full context while adapting tone and framing for each language.

Agent Mode vs Script Mode

I used both Claude Code's agent mode and raw shell scripting. Each serves a different phase:

Agent mode for prototyping:

"Build an AI news generation script. Requirements: Claude API calls for 4 articles, English/Korean simultaneous generation, auto-frontmatter, auto-filename from date and title."

Agent mode lets you refine requirements conversationally, modify multiple files simultaneously, and discover edge cases early.

Shell scripts for production: The actual daily pipeline is pure bash — generate-ai-news.sh with direct curl calls to Claude API. No interactive session needed. The script runs on a cron, commits results, and deploys.

MCP for Real-Time Data

News quality improves with real-time source data. MCP (Model Context Protocol) servers feed Claude live information:

{
  "servers": {
    "news-feeds": {
      "command": "node",
      "args": ["mcp-servers/news-rss.js"],
      "env": { "FEEDS": "techcrunch,theverge,wired,arstechnica" }
    }
  }
}

With MCP, Claude references current headlines instead of relying on training data. The generated content is more accurate and timely.

CLAUDE.md for Quality Consistency

Daily content generation demands consistency. CLAUDE.md locks down the quality standard:

## Content Standards
- Fact-check all claims with multiple sources
- Include specific numbers and dates
- Avoid speculation unless clearly marked

## English Version
- Professional, concise tone
- Business impact focus
- 300-500 words per article

## Korean Version
- Developer-friendly tone
- Use 반말 consistently
- Include technical implementation details
- 400-600 words per article

## File Naming
- Format: YYYY-MM-DD-{key-topic}-{company/tech}.md
- kebab-case, max 60 characters

With this in place, Claude produces consistent quality regardless of session.

Platform-Specific Frontmatter

Same content, different metadata requirements:

jidonglab.com (Astro):

title: "Nvidia Announces Rubin GPU Architecture at GTC 2026"
date: 2026-03-14
category: "ai-hardware"
tags: ["nvidia", "gpu", "gtc", "rubin"]

DEV.to:

title: "Nvidia GTC 2026에서 Rubin GPU 아키텍처 공개"
published: true
tags: ai, gpu, nvidia, hardware
canonical_url: https://jidonglab.com/ai-news/2026-03-14-nvidia-gtc-rubin
series: "AI 뉴스"

A convert_frontmatter() function handles the transformation between formats. The GitHub Actions workflow runs the generation script, converts frontmatter, publishes to DEV.to, commits, and deploys — all on a daily cron.

What Could Be Better

Tool Use for real-time fact-checking. Claude's tool use feature could verify company information, funding rounds, and source accuracy during generation rather than after.

Vector DB for semantic deduplication. Keyword-based dedup misses articles that use different words for the same story. Embedding-based cosine similarity would catch those.

Two-stage generation for cost optimization. Use Claude 3.5 Haiku for first drafts (cheap), then Claude 3.5 Sonnet for quality review (accurate). Daily generation adds up, and this pipeline cuts costs significantly.

Takeaways

• Constraint-heavy prompts produce consistent quality; open-ended prompts produce variance
• MCP server integration gives Claude access to real-time data for more accurate content
• Template-based structure is more reliable than free-form generation
• Full automation means designing the entire workflow — generation, formatting, validation, deployment — not just the writing step

Commit log

e615288 — feat: AI news script — English (jidonglab) + Korean (DEV.to) dual generation
a00b3bf — feat: AI news 2026-03-14 (4 posts, en)
069ca0d — feat: AI news CLI generation script + 2026-03-14 4 news items
6788360 — feat: AI news auto-generation (2026-03-14)