Claude-Powered News Pipeline: From API Call to Production Markdown in One Script

Running an AI news site means generating content every single day. I automated the entire workflow: one command produces 4 validated, formatted, publication-ready markdown files. Claude API + shell script + MCP file access + git hooks. Here is how each piece fits together.

The One-Command Target

The goal was simple: type "generate 4 AI news for 2026-03-14" and get finished markdown files committed to the repo. Not ChatGPT-quality drafts that need editing — production-ready content with correct frontmatter, validated structure, and consistent tone.

Structured News Generation Prompts

The prompt that produces publishable content:

"Write 4 AI news articles in Korean for 2026-03-14. Each article:

1. Title: 40-60 characters, keyword-rich, no clickbait
2. Body: 3-4 paragraphs, lead with the key fact, close with outlook
3. Tags: ['AI', 'News'] required + 2-3 company/tech names
4. Filename: date-keyword-2-3-words.md format

Do not cover: rumors, speculation, stock-related content
Focus on: technical advances, policy changes, product launches, research results"

I chose Claude over GPT-4 and Gemini for this task because of its stronger instruction following on long, structured output. When the prompt says "3-4 paragraphs" and "40-60 characters," Claude delivers within bounds more reliably.

The failing prompt: "Write some AI news." No length constraint, no structure, no source filtering. Every run produces different quality.

The Shell Script Pipeline

generate-ai-news.sh is not just an API wrapper. It orchestrates the full content pipeline:

#!/bin/bash
DATE=${1:-$(date +%Y-%m-%d)}
COUNT=${2:-4}
LANG=${3:-ko}

response=$(curl -s -X POST https://api.anthropic.com/v1/messages \
  -H "Content-Type: application/json" \
  -H "x-api-key: $ANTHROPIC_API_KEY" \
  --data "{
    \"model\": \"claude-3-sonnet-20240229\",
    \"max_tokens\": 4000,
    \"messages\": [{\"role\": \"user\", \"content\": \"$PROMPT\"}]
  }")

echo "$response" | jq -r '.content[0].text' | split_and_save_files "$DATE"

The split_and_save_files function parses Claude's response (4 articles in one response), extracts keywords for filenames, and saves each article as a separate file.

Bilingual Generation in a Single API Call

Instead of two separate API calls for English and Korean, I bundled both in one request:

"Write 4 news articles. For each, provide both Korean and English versions. Do not translate — write each version for its target audience with appropriate tone and examples.

Output format:

News 1 (Korean)

[content]

News 1 (English)

[content]"

One API call, two languages, consistent pairing. Translation quality is significantly better than running separate calls because Claude maintains context across both versions.

MCP for Direct File System Access

Claude Code's MCP (Model Context Protocol) enabled direct file system access, eliminating the copy-paste workflow:

{
  "mcpServers": {
    "filesystem": {
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-filesystem"],
      "env": {
        "ALLOWED_DIRECTORIES": ["/Users/jidong/projects/jidonglab.com/src/content"]
      }
    }
  }
}

With MCP, Claude reads existing articles (for dedup checking), writes new files directly, and updates metadata across 20+ posts in a single batch operation. For the batch metadata update:

"In src/content/blog/, across all .md files: rename publishedDate to publishDate, deduplicate tags arrays, truncate description to 150 characters. Show the change plan before applying."

The "show before applying" clause lets me review 20 file changes at once instead of editing one by one.

Content Quality Validation

AI-generated content has inconsistent quality. Three validation strategies brought it under control:

Self-check in the prompt:

"After writing each article, verify: does the title match the content? Does the lead paragraph contain the key fact? Are sources and dates accurate? Are technical terms explained? Revise if any check fails."

Template enforcement:

---
title: "Title (40-60 chars)"
description: "Summary (100-150 chars)"
publishDate: "YYYY-MM-DD"
tags: ["AI", "News", "keyword1", "keyword2"]
---

## Key Summary (2-3 sentences)
## Main Content (bullet points with specifics)
## Industry Impact
## Related Links

Forcing a template reduces omissions. It also makes later human editing faster because the structure is predictable.

Context-based deduplication:

"Recent topics covered: Anthropic Claude, NVIDIA H100, OpenAI GPT-5, Google Gemini... Write new stories that do not overlap with these topics."

Cannot fit full articles in context, but keywords and summaries are enough to prevent duplicates.

Git Hooks for Automated Commit

The script does not stop at file creation. It commits and optionally deploys:

git add src/content/ai-news/
git commit -m "feat: AI news $DATE ($COUNT posts, $LANG)"

if [[ "$DEPLOY" == "true" ]]; then
  npm run build && npm run deploy
fi

A pre-commit hook runs markdown lint on generated content. AI-authored content meets the same quality bar as human-authored content.

What Could Be Better

RSS feed monitoring for autonomous topic selection. Currently I specify "generate 4 news items" manually. Parsing RSS feeds from major tech outlets would make the system fully autonomous.

Vector DB deduplication. Keyword matching misses semantic duplicates. Embedding articles and computing cosine similarity against the existing corpus would catch stories that use different words for the same event.

Claude tool use for live fact-checking. Instead of post-generation validation, tool use functions could verify company data and source URLs during generation.

Analytics feedback loop. Feeding Google Analytics data (page views, time on page, shares) back to Claude as context would let the system learn which article styles perform best.

Takeaways

• Specific constraints in prompts (character count, structure, source filtering) produce consistent quality
• MCP file system access lets Claude manage content directly — no copy-paste intermediary
• Template-based structure is more reliable than free-form generation for daily content
• Automating the full workflow (generation → validation → commit → deploy) eliminates the human bottleneck

Commit log

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)