Writing with AI sucks
Why handing raw notes to an LLM loses alignment and intent, and the two-skill workflow I use so the text stays mine.
Contents
In the past few months, I have tried to simplify and automate blog posting. During this process, I have created all kinds of tests:
- I created a fully automated workflow where, with a simple prompt, AI researched a topic and created everything, including the diagrams and tables.
- I created a semi-automated workflow where I used the
/grill-meskill to figure out the content, outline it, and draft it, and eventually an LLM such as Opus 5 would write the final article (.mdx).
All of this is because I wanted to save time. For me, the blog is a way to show my interests, perhaps share interesting information, and store my public notes for later.
But the issue is that LLMs are just really bad at writing something more complex than a simple email. I am not a writer, but I have already crafted a lot of text. I don’t really care about writing style; I just want to share some of my ideas.
However, the issue I found with letting an AI write even a simple text is that alignment with my intent is almost impossible to achieve. Why? Because notes and dictation as raw material are not the same as writing and editing an article. It is not the same mental care. If the raw text were already what I wanted, I would not need an LLM to write it.
When you hand only raw notes to an LLM—even the latest frontier model—it bridges those islands of constructs through generalization, causing the text to lose its point, tone, mistakes, and, essentially, its value. When you edit the draft yourself, you put that mental work into the text, and the LLM acts like a tool rather than an author.
The workflow I landed on
So I will follow what others do: the text will be mine, and I have removed my old AI-written posts. Instead of asking an LLM to write the article, I now use a two-skill workflow:
draft-shaper: takes my draft, organizes my verbatim text into sections, and inserts[TODO]questions directly into the draft where my thinking or bridges are missing. I write 100% of the words, and the AI only fixes typos, adds external links, and checks if the argument holds up.blog-packager: takes the finished Markdown draft and handles the frontmatter, assets, HTML/Astro Islands elements, and.mdxvalidation.
This approach seems to work better because even the corrections are coming from me, not the LLM. The LLM just adds signals, which I act on. This might also help to improve my writing skills.