Building an Auto Dev Blog Generator

Wanted a way to document my dev work without spending hours writing blog posts. Ended up building an automated generator using Claude Code skills.

The problem

After coding sessions, I’d have useful knowledge to share but writing it up felt like a chore. Plus when I did write, it either sounded too robotic (AI-generated vibes) or took forever to make it natural.

Needed something that could:

• Turn conversations into posts automatically
• Sound like I actually wrote it
• Work for study groups and helping others learn
• Be in English for wider reach

What I built

Created a Claude Code skill that analyzes conversations and generates blog posts. It’s a slash command that does everything:

/post

That’s it. It reads the conversation, extracts what matters, writes a post, and pushes to GitHub.

How it works

The skill lives in my dotfiles:

~/.claude/skills/write-devlog/
├── SKILL.md       # Main logic
└── config.json    # Blog repo settings

Config setup

Points to my Astro blog repo:

{
  "blog": {
    "repo_path": "/Users/direcision/Workspace/ww/restato.github.io",
    "posts_dir": "src/content/blog",
    "remote": "origin",
    "branch": "main"
  },
  "defaults": {
    "language": "en",
    "writing_style": "casual_personal",
    "hide_ai_traces": true
  }
}

The skill logic

SKILL.md tells Claude how to:

1. Analyze conversations - Pull out code snippets, problems solved, concepts learned
2. Generate natural text - Use phrases like “I tried”, “turns out”, “spent 2 hours on this”
3. Variable structure - Not every post needs the same sections
4. Remove AI hints - No “Co-Authored-By”, no perfect formatting
5. Auto-publish - Create MDX, commit, push to GitHub

Key part was making it sound human. The skill uses casual language and includes struggle:

# Good examples from the skill:
- "Ran into this error..."
- "Spent 2 hours debugging before realizing..."
- "Got frustrated with X"
- "Here's what worked"

Things to know

You’ll need:

• Astro blog (or adapt for other static sites)
• Claude Code CLI
• Blog repo locally cloned
• Git configured

The skill generates MDX with frontmatter:

---
title: "Generated Title"
description: "One sentence summary"
date: 2026-01-13
tags: ["tech", "tags"]
---

Gotchas

AI detection was the big one. First version sounded too perfect. Had to explicitly tell the skill to:

• Vary post structure
• Use conversational transitions
• Show personality and struggle
• Allow minor imperfections
• Never add “Co-Authored-By” or similar

Also realized every post having the exact same sections screams AI. Now it picks sections based on content - sometimes includes “Prerequisites”, sometimes “Gotchas”, sometimes neither.

File naming - Convert titles to kebab-case, handle collisions with timestamps:

"Building Auto Generator" → "building-auto-generator.mdx"
// If exists: "building-auto-generator-1432.mdx"

What it generates

Posts include:

• Natural intro (what you did and why)
• Core “What I did” section
• Optional sections: Quick summary, Prerequisites, Gotchas, Links
• Casual outro when it makes sense

All in English, all sounding like a person wrote it while coding.

The workflow

# Work on something
# Have conversation with Claude
# When done:
/post

# Output:
# - Analyzes conversation
# - Generates post
# - Commits: "Building Auto Generator"
# - Pushes to main
# - Shows GitHub + live URLs

Takes about 30 seconds, then it’s live.

Why this approach

Tried a few things before this:

1. Manual writing - Too slow, often didn’t do it
2. AI with heavy editing - Still took forever to make it natural
3. Templates - Everything looked the same (AI vibes)

This works because it generates variable content that adapts to each conversation. No two posts are identical in structure.

Links

• Claude Code Skills docs
• My blog repo

Next steps

Might add:

• Preview mode before publishing
• Series support (Part 1/2/3)
• Difficulty tags (beginner/advanced)

But keeping it simple for now. It works, posts get published, people can learn from them.