Using AI to help write articles
How I use AI in the loop when writing articles for this site: scaffold first, rewrite everything, bank the corrections, and never publish what you don't know.
Most of the recent articles on this site were written with an AI in the loop, and it seems only fair to be upfront about how I use AI tooling for my writing. If you're starting to use these tools for your own writing, or you've tried them and everything comes out sounding like a press release [1], the process I've been working through might be useful. In my experience AI works well as an efficiency tool and badly as a content creator, and everything below follows from that.
Start with a scaffold
Before the age of AI I used to plan articles with a scaffold first - decide what the piece is for, who it's for, what the headings are and roughly how long it should run, and only then start writing. It's mostly a cure for the blank page: a handful of bullet points turns "write an article" into "fill in the gaps", and the writing starts moving. The scaffold isn't fixed, though. It nearly always changes as the draft takes shape, and the real rhythm is a loop - scaffold, draft, restructure, write some more, repeat.
The only thing that's changed is that the scaffold is now agreed with an AI before any drafting starts, and the raw material is still mine: a rough draft I've already written, or at the very least a long prompt covering what I've been thinking about, the angle I want to take and the experience it draws on. Ask a model to "write me an article about serverless search" with only that to go on and it will cheerfully produce a thousand words of confident, generic filler, because it's filling the space with what everyone else has already written. Give it the thinking, an audience and an ordered set of headings, and the draft that comes back is something worth editing. I also keep the drafting incremental - a section at a time, reacting as it goes - rather than taking delivery of a finished essay.
Rewrite it in your own words
The draft is a starting point, and the rewriting is where the article actually gets written. I go through and add detail, change phrases I'd never say and cut sentences that are too pleased with themselves, and rework whole paragraphs by hand when the shape is wrong. There are usually several rounds of this, and the article that gets published can be a long way from the draft that started it.
One rule I have is never writing about things I don't know - which frustratingly AI makes very easy to break because it will happily produce an authoritative paragraph on any subject. While writing the search series it drafted a section on database storage internals with a fluency I simply don't have, and it read like someone else entirely. The test I use is whether I could stand up and talk about the subject unprompted. If I couldn't, the section either goes, or gets rewritten down to the depth I actually have, curiosity and open questions included. Passing off machine expertise as your own is the quickest way to make writing feel fake.
Teach it as you go
Every correction I make gets banked in a writing style file the AI reads before it writes anything, so when a draft comes back with something that isn't me, the fix becomes a permanent rule rather than a one-off edit. This file started as an analysis of stuff I had written over the years - AI is quite good at reviewing material and summarising how the writer writes. Then over a few months of use and iteration the file has grown into a fairly good description of how I actually write - the phrases to avoid and the rhythm of the sentences. It works far better than telling a model to "write like me", which gets you a caricature.
If you're starting with your own AI writing then building a style guide is the single best first step I'd recommend. Correcting the same classic AI phrases in every draft is a pain in the arse, but a correction saved somewhere persistent means the tool comes back a little more like you each time. That compounding effect is really powerful and is one of the things I love about AI tooling. I use a GitHub repo to store my style guides and skills - but any cloud storage synced to your computer would work as well.
The proofread
Finally I read the "finished" piece through, as a reader rather than the writer, and ask whether it's actually interesting. If it's boring, too wordy or has nothing to say, it gets edited again or binned, and not everything survives that read. Facts get checked against sources before anything goes out too, because a language model will state a plausible wrong thing with complete confidence. I have a much bigger pile of rejected ideas than ever before because AI helps to get things written down to a point where you can judge if it's worth carrying on with, rather than staying an idea that never goes anywhere.
So the ideas, the opinions and the mistakes in these articles are mine. The AI's job is to get me to a publishable draft in hours rather than days (or more likely not at all). It's an amazing efficiency tool but I don't want it to become the author.
[1]: AI does do a fantastic job of writing press releases though - that's a use-case where it can lean on so many existing real-world examples and be fed with facts, and where the voice is deliberately not a person.