Essays • AI and engineering • • 5 min read

AI Development Revolution Part 4: Content Pipeline

Turning years of handwriting and a tangled WordPress blog into something I could maintain, without letting the machine flatten the voice out of it. OCR for cursive, a WordPress to Astro migration, and where automation actually ends.

Illustration of a handwritten page passing through OCR and AI markers into stacked cards labeled Articles and Blog Posts, with a 363+ Articles badge

The Content Pipeline

Part 4 of the AI Development Revolution. This was the long one, and most of what I learned later started here.

The problem was my own handwriting

Daily freewriting is my morning practice. Two to five pages of cursive before the day starts. The writing was never the problem. Getting it off the page was.

Generic transcription models could not read me. My handwriting is its own dialect, and the early models guessed at it badly enough that fixing their output took longer than just typing the pages myself. But I write a lot, and retyping all of it by hand was not going to last. So I needed something in between, a pipeline that could take a photo of a page and hand back a clean draft that still sounded like me.

I had written before, in "Taming the Paper Tiger", about how thin the thread is between thought and text. Speeding it up with a machine was the fastest way to snap it.

Five tries before it worked

The first version used basic OCR. It read the words and lost the voice completely, flattened everything into the same gray tone. The second used a better model, and somehow that was worse. It cleaned the writing up into corporate copy, the kind of prose that says nothing in a confident way. I tuned the prompts and configured a tool around them, and it came out too rigid, locking onto patterns and forcing every page through them. I tried a context-aware pass that was supposed to read the meaning and keep the style. It fell apart on stream-of-consciousness, which is most of what freewriting is.

It did work in the end, after complete rewrites and weeks of debugging and prompt tuning, with most of it thrown out along the way.

The working version came from doing it badly four times and paying attention to how each one failed, with no clever idea anywhere in it. It cost weeks of work scrapped, the same pages retyped to check the machine against the truth, and the stubbornness to start over with a different approach when the last one had felt close.

I wrote in "Handwriting as Meditation" and "The Flow of Thoughts" that the link between the hand and the thought is the point. Getting a model to leave that link alone was the whole job.

The WordPress exodus

The other half was years of old posts trapped in a WordPress database. I wanted them out, in plain markdown files I could read and move and back up without a plugin and a server in the way.

The export was easy and the cleanup was not. I ran the first hundred articles through an enhancement pass and they all came back sounding like a brochure, smoother and shorter and gone. So I read all hundred with a red pen, rebuilt the rules, and read them again. The "automation" moved the work instead of removing it. Instead of writing the articles I was reading every one of them, slowly, asking whether the machine had quietly swapped a sentence of mine for a sentence of its own.

A few things it was genuinely good at, and I let it have them. Alt text for images. Formatting code blocks. Checking links. Writing a meta description that summarized a post without trying to improve it. Anywhere it had to be descriptive instead of creative it earned its keep, and everywhere it tried to be me I took the work back.

What I gave up, and what I kept

Every step was the same trade. Making a post clearer risked sanding off the part that was interesting, and optimizing a title for search meant losing the title that had a joke in it. Most of the time I chose the joke. I left original titles when they had character and accepted that they would rank worse for it. I kept transitions that were a little rough, because rough was mine and smooth was anyone's.

The pipeline only started working when I stopped trusting it. Scanned page in, draft out, and then a person reading the draft against the page before any of it went near the blog. That last step never went away, and at some point I stopped trying to make it.

What it is for

What I have now is quiet. A photo of a morning's pages becomes a draft by the time I have had coffee, and the draft still sounds like the person who wrote it longhand. The old posts are files I own instead of rows in a database I rent. None of it is hands-off, and I have made peace with that.

I wrote about working with the machine as a partner instead of a tool in "Vibe Coding with AI". This project is where I learned where the partnership ends, which is at the page, with me reading.

Next I took the same habit, build it badly and watch how it breaks and keep a human in the loop, and pointed it at whole applications, which is Part 5.

Next: Part 5: The Business Transformation →

Series Navigation:

AI development