Aadil Mamujee

I’ve been writing on my personal blog since 2012 (122 posts over 14 years) and many of them are decent, but could also have been written by anyone.

I want to be a better writer. I’ve read books, hired an editor, and asked friends for honest feedback. But it’s hard to be objective about your own work, and even harder to know if you’re improving. I’m not optimizing for distribution or shareability so views/shares are not the right metric. I care about the quality of thinking, documenting the things I’m passionate about, and sounding like myself when I do it.

My best posts – like the neuromuscular condition one or turning forty are my most personal. The weakest ones usually started with a framework from a book I’d just read or were structured like product memos.

I exported all my blog posts into a Claude project and got it to tell me what I did well and what my issues were. It packaged the findings into a custom skill, which is a set of instruction that loads whenever I start a new draft. Now there’s a partner in my publishing workflow that has read everything I’ve ever published and compares the new writing to the old.

I tried to dictate rough drafts into Superwhisper, structure with Claude, then run the draft through multiple other models for feedback. The feedback is usually the same: go deeper into the things only I would know.

The obvious conclusion is that the writing quality has improved and this writing system works. I’m not convinced. There’s a real risk that what feels like coaching is the AI quietly pulling me toward what all the AIs think good writing looks like. Reversion to the mean, dressed up as feedback. Sentences get shorter (example intended). Hedges get cleaned up, but so do the sentences that are awkward on purpose and sound like something I would actually say.

My wife read a few heavily AI-edited drafts and told me they sounded like an LLM wrote them, not me. Since then I’ve been second-guessing almost every edit that the llms propose. I should be the writer and the editor, not the model. But the model might also be right, and I don’t always know when I’m losing my voice versus making the writing better.

I don’t have this figured out. The writing system is good at catching things that are genuinely bad: hedging, bullet points that should have been sentences, generic frameworks. It’s less good at knowing when a sentence is awkward on purpose, when a tangent is the point, when the worst-flowing paragraph in a post is the one I most need to keep. That’s where I have to push back on the edits – sometimes I will, and other times I’ll cave and regret it later. There will be posts that come out of this process a little too “clean”, and I’ll know which ones when I reread them in a year.

I’d rather find out a draft is mediocre while I’m still editing it than six months after I published it, and the AI supported writing system gets me there faster. The cost is that the draft I’m editing is already, slightly, not mine. The version that gets published is the version that goes through the smoothing, and that version is closer to what the LLMs think is good writing – reversion to the mean.

For now, I’m treating the writing system as a reader I trust when I have a fully written draft by me, a human. My voice is the thing I bring and the writing system is there to tell me when I’ve drifted away.