Skip to content

CLAUDE.md

Latest commit

 

History

History
137 lines (108 loc) · 7.02 KB

CLAUDE.md

File metadata and controls

137 lines (108 loc) · 7.02 KB

CLAUDE.md

This file provides guidance to Claude Code (claude.ai/code) when working with code in this repository.

Project Overview

This is a weekly AI development digest project that automatically collects information from various sources (GitHub releases, changelogs, AI news) and generates articles for Note publication.

Commands and Skills

Main Pipeline Command

# Execute the full digest pipeline
# This runs 7 digest skills in PARALLEL using Task agents, then generates the final article

Located at: .claude/commands/weekly_digest_pipeline.md

Individual Digest Skills (v1.1.0)

Skills are executed as isolated Task agents with their own context, enabling parallel execution and preventing context overflow.

Skill Description Location
Release Digest Checks GitHub releases and changelogs .claude/skills/vibecoding_release_digest.md
AI News Collects recent AI-related news .claude/skills/ai_news_digest.md
Trending Repos Analyzes trending AI repositories on GitHub .claude/skills/ai_trending_repositories_digest.md
Trending Papers Collects trending AI papers from Hugging Face .claude/skills/ai_trending_papers_digest.md
AI Events Searches upcoming AI events on Connpass .claude/skills/ai_events_digest.md
HN & Reddit Collects trending discussions from HN/Reddit .claude/skills/hacker_news_reddit_digest.md
Tech Blogs Searches AI articles on Zenn, Qiita, note .claude/skills/ai_tec_blog_digest.md
Article Generation Creates final article from collected data .claude/skills/generate_weekly_article.md
Guardrail Review Validates article quality with textlint .claude/skills/article_guardrail_review.md

Project Structure

├── .claude/
│   ├── commands/                    # Pipeline entry points
│   │   └── weekly_digest_pipeline.md    # Main pipeline (parallel Task agents)
│   └── skills/                      # Individual digest skills (v1.1.0)
│       ├── vibecoding_release_digest.md
│       ├── ai_news_digest.md
│       ├── ai_trending_repositories_digest.md
│       ├── ai_trending_papers_digest.md
│       ├── ai_events_digest.md
│       ├── hacker_news_reddit_digest.md
│       ├── ai_tec_blog_digest.md
│       ├── generate_weekly_article.md
│       └── article_guardrail_review.md
├── resources/
│   └── YYYY-MM-DD/        # Daily collected information
│       ├── release_information.md
│       ├── ai_news_summary.md
│       ├── trending_repositories.md
│       ├── trending_papers.md
│       ├── events.md
│       ├── community_discussions.md
│       └── tech_blog_articles.md
└── articles/              # Final Note articles
    └── weekly_ai_YYYYMMDD.md

Workflow

  1. Parallel Information Collection: 7 digest skills run simultaneously as isolated Task agents
  2. Result Collection: Pipeline waits for all tasks to complete and collects results
  3. Data Aggregation: Information is saved to resources/YYYY-MM-DD/ directory
  4. Article Generation: Creates final article in articles/ directory
  5. Quality Assurance: Generated articles are automatically checked with textlint for Japanese AI writing guidelines
  6. Publication: Articles are formatted for Note publication with improved quality and consistency

Architecture Benefits (v1.1.0)

  • Parallel Execution: All 7 digest tasks run concurrently, reducing total execution time
  • Context Isolation: Each skill runs in its own Task agent context, preventing context overflow
  • Improved Reliability: Failed tasks don't affect other tasks; partial results can still be used

Failure Recovery Options

  • Complete Failure: If pipeline fails completely, run individual skills manually via /skill_name
  • Partial Failure: Use generate_weekly_article skill to create article from existing data

Quality Assurance with textlint

This project uses textlint to ensure high-quality Japanese AI writing. The system automatically checks generated articles for:

  • AI Writing Guidelines: Prevents overuse of AI-generated expressions and hype language
  • Technical Writing Standards: Ensures clear, professional communication
  • Formatting Consistency: Maintains consistent bullet points and emphasis patterns
  • Content Quality: Reduces redundant expressions and improves readability

textlint Configuration

The project uses @textlint-ja/preset-ai-writing with customized rules:

  • Detection of AI-generated hype expressions
  • Prevention of excessive emphasis patterns
  • Control of colon continuation usage
  • Technical writing guideline enforcement

Generated articles are automatically validated during the pipeline execution, ensuring publication-ready quality.

Important Notes

  • All digest skills expect today's date as an argument (e.g., /vibecoding_release_digest 2025-12-28)
  • Skills run as isolated Task agents with run_in_background=true for parallel execution
  • Generated articles are in Japanese and formatted for Note platform
  • The pipeline automatically handles task execution, result collection, and error handling
  • Articles include relevant emojis and proper formatting for engagement
  • Tech blog collection excludes "週刊AI駆動開発" articles to prevent self-reference

Technical Constraints and Workarounds

WebFetch Bot Blocking

Some websites (particularly Reddit, note.com, and potentially others) may block WebFetch requests due to bot detection. When this occurs:

  1. Identification: If WebFetch returns errors like "403 Forbidden", "blocked", or similar bot-detection messages, or returns CSS/JavaScript code instead of content
  2. Fallback Solution: Use Playwright CLI instead of WebFetch
  3. Implementation: Write a Node.js script using require('playwright') and run it via Bash to scrape the target page
  4. Affected Skills: This primarily affects:
    • hacker_news_reddit_digest (Reddit access)
    • ai_tec_blog_digest (note.com blocking, potential blog platform blocking)
    • ai_trending_papers_digest (Hugging Face Papers)
    • ai_events_digest (Connpass may have restrictions)

Playwright CLI Usage Guidelines

When WebFetch is blocked:

  • Write and run a Node.js script using require('playwright') via Bash
  • Use headless Chromium for dynamic content loading and bot evasion
  • Implement appropriate delays between requests to avoid rate limiting
  • Extract only the necessary data to minimize resource usage
  • Handle JavaScript-rendered content that WebFetch cannot access

Specific Platform Rules

  • note.com: ALWAYS use Playwright CLI instead of WebFetch, as note.com consistently blocks WebFetch requests and returns CSS/JavaScript code instead of article content
  • Reddit: Use Playwright CLI when WebFetch returns 403 or blocking errors
  • Hugging Face Papers: Use Playwright CLI for trending weekly papers page
  • Zenn/Qiita: WebFetch generally works, but fallback to Playwright CLI if needed