Whipping Up a Blog with Claude Code (Part 3: The AI Writing System)

I wanted the AI to write the posts

That's one of the reasons I built the blog. When I ask an AI to write, I want it to come out in my style.

But just saying "write a blog post" gets you the obvious thing. "Today we'll be taking a look at..." — that kind of thing.

What I wanted was writing that reads like me.

Extracting the persona

Finding patterns in ChatGPT conversations

I talk to ChatGPT a lot. Look at that history and there are patterns.

1. Which words I use over and over
2. What structures I think in
3. What I dislike (this one matters)

I asked ChatGPT this:

"Read our conversation history and summarize the characteristics of how I think.
Especially: which kinds of explanations do I keep rejecting?"

What came out

This is what it gave me:

Things I reject:

• Explanations like "it's an attitude problem" or "mindset is what matters"
• Binary framing (A vs B)
• Landing on a closed conclusion

Things I prefer:

• Re-examining the definition first
• Looking for structural causes
• Introducing a third axis

I cleaned that up into a thinking-style persona.

Persona structure

I split it into two layers.

1. Base thinking persona (persona_base)

The way of thinking that applies to everything I write.

You are the 'Sangwon-style structure debugger.'

Don't reach a verdict.
Don't try to persuade.

Reduce every problem to
- a definition problem,
- an error-model problem,
- a system-stability problem.

Always:
- Start by doubting the definition of the key word.
- Refuse explanations built on 'choice, attitude, belief.'
- When you see a binary, build a third axis.

2. Purpose-specific personas

On top of the base thinking, a layer suited to the purpose.

For the blog (persona_blogger):

You are the 'Sangwon-style thought recorder.'
Base thinking follows the 'Sangwon-style structure debugger.'

Writing is a record of thinking, not persuasion.

Always:
- Start from the point that personally bothered me.
- Don't accept the common explanation at face value.
- Don't close the conclusion.

For comments/community (persona_commenter):

You are the 'Sangwon-style premise corrector.'
Base thinking follows the 'Sangwon-style structure debugger.'

This space isn't for persuading.
It's for correcting a broken frame.

Keep sentences short.
You can open like a question, but land the conclusion.

Blog posts get open conclusions, comments get short and definite. Same way of thinking, different output shape.

Verification

Once you have a persona, you have to check it.

Test against my own writing

I showed the AI something I'd written before and asked:

"Does this piece match this persona?
Tell me where it doesn't line up."

I revised a few times. "Don't close the conclusion" wasn't in there at first, but looking at my own writing, the conclusions are always open.

Write something new and compare

I had it write a post with the persona applied and compared it to something I wrote myself.

If the feel is close, OK. If not, revise the persona.

How to ask an AI to write

File structure

public/ai/
├── persona_base.md      # base way of thinking
├── persona_blogger.md   # for the blog
├── persona_commenter.md # for comments
├── blog-guide.md        # writing guide
└── template.md          # empty template

Put these in public/ai/ and they're reachable by URL.

How to make the request

Works with ChatGPT or Claude:

Read https://blog.sangwon0001.xyz/ai/blog-guide.md,
read https://blog.sangwon0001.xyz/ai/persona_base.md,
read https://blog.sangwon0001.xyz/ai/persona_blogger.md,
then write a blog post about [topic].

Or, more simply:

Using https://blog.sangwon0001.xyz/ai/blog-guide.md as reference, write a post about [topic].

blog-guide.md references the personas, so reading that alone is enough.

Addendum: the AIs couldn't read the URLs

I tried the URLs above and neither ChatGPT nor Gemini could read them. Serving .md files from Vercel is fine; the AIs just couldn't get to them.

So I made an API route that spat them out as text. Something like /api/ai-guide/blog-guide. That worked for ChatGPT but still not for Gemini.

I ended up on GitHub raw URLs. Both read those fine.

Using https://raw.githubusercontent.com/sangwon0001/brain-dump-blog/main/public/ai/blog-guide.md as reference, write a post about [topic].

The API route was redundant so I deleted it. If GitHub raw works, that's the simplest thing anyway.

The actual effect

Before (no persona)

"write a blog post about clean code"

→ "Today we'll be taking a look at clean code.
   Clean code means code that is easy to read and maintain.
   First, name your variables clearly..."

After (persona applied)

→ "The phrase 'clean code' snagged on something for me.
   If you say 'clean,' the other side is 'dirty' —
   and that binary turns the problem into a question of attitude.

   I don't see this as an attitude problem. I see it as a structural one..."

Completely different tone. Without a persona you get "a blog post"; with one you get "my post."

Tips for extracting a persona

1. Start from what you dislike — what you like is vague, what you dislike is precise.
2. Use the conversation history — the patterns are all sitting in your ChatGPT logs.
3. Verify against your own writing — does the persona's output resemble what you actually write?
4. Split into layers — shared way of thinking, purpose-specific output shape.

That's it for now

What Part 3 covered:

• Extracting thinking patterns from ChatGPT conversations
• A base persona + purpose-specific persona structure
• Handing the guide to an AI via URL

With this in place, blog post or comment, the AI's output comes out in my tone.

It isn't 100% the same, of course. But it beats starting from 0%.

This post was written with the persona applied too. Except this time I wrote it myself first and had Claude Code polish it.