When I started this blog, one of the things I found interesting was that I had used coding agents to make it. I found a blog I liked, cloned it, used AI to help me reproduce the bones of it, then gradually changed it into something of my own.
Great, the site exists, but a few months later the website itself is no longer the bit I find most interesting. The system around it has become much more elaborate than the thing I first set out to build. There is now an EDITORIAL.md, an AUTOMATION.md, an AGENTS.md, a WORKFLOW.md and, as of this week, a VOICE.md. GitHub Issues hold half-formed ideas. Pull requests record what changed and why. There are checks for provenance, metadata and whether the site still builds. More recently, the system has started recording editorial lessons from the places where generated writing did not feel quite like me. For a personal blog, that is probably a lot of over-engineering!
Working through recent posts has made the gap between readable and recognisably mine easier to see. A generated draft can be perfectly readable and still feel wrong. A small observation gets turned into a framework because the system thinks an essay ought to arrive somewhere. A trip becomes a theory. A human detail gets explained for another paragraph after it has already done the work. An ending becomes suspiciously neat. Sometimes it is even smaller than that: one draft used the word ephemeral. Perfectly good word. I would never use it.
Those corrections are now starting to become part of the system. We have written down things like: do not turn a small observation into a framework; concrete detail can carry more than another theory paragraph; author-supplied language outranks a smoother synonym. They are actual editorial learnings in the repository.
That changes the experiment slightly. The system is no longer only being told to write in a style that looks a bit like mine. It is accumulating evidence about the editorial decisions I make when something doesn't feel right. Style is relatively easy to copy. You can count sentence lengths, spot recurring words, copy the rhythm of a few published posts and probably produce something convincing at a glance. The harder bit is deciding that a claim has gone too far, a paragraph is trying too hard to sound profound, the odd little detail is more important than the explanation around it, or something should stay unresolved rather than being converted into a conclusion because conclusions are neat and tidy.
Those decisions are not especially dramatic, and most would never be visible to a reader because the rejected versions disappear. Now some of them are becoming part of the system. The risk is that, in learning what I preferred yesterday, it becomes less able to recognise when I want something different today.
A preference that was useful across five articles might be exactly the wrong instinct for the sixth. I may change my mind or deliberately want to write something differently. We have dealt with that, at least in the design, by making the hierarchy explicit: what I actually say or change in the current piece outranks anything the system thinks it has learned about me previously. Otherwise a personalised system could become a very efficient way of being trapped by your own history.
I noticed a version of that problem yesterday when I read one of my older posts and noticed how much more "AI'd" it seemed than my recent work. I want to keep open the option of rewriting pieces like that when the residue becomes distracting, without turning the archive into a permanent polishing exercise or pretending I had today's thoughts and tools when I wrote something months ago.
Which makes the timing of this experiment fairly amusing (to me at least). This week former Anthropic researcher Jacob Coxon resigned while warning about the direction of the AI race, and Anthropic's Alignment Science Lead Evan Hubinger said he thinks there is a greater than 10% chance AI could kill all humans within the next decade. Meanwhile, I am using the technology to work out how much of "me" I can sensibly delegate to a publishing system. The stakes are, admittedly, slightly different.
There is still something useful in the contrast. I do not want the system to decide what Phil sounds like and then become very good at passing its own test. If it is going to carry more of the mechanics, I want the current version of me to remain able to contradict what it learned from the previous one.
Does that make the writing more mine? I don't know yet. It might simply move my effort downstream, from writing every sentence to noticing what the system misunderstood. There is still a reasonable question about whether that is authorship, editing or something 'new' sitting between the two.
This article is the first one being put through that loop deliberately. We kept the first draft and the critique. I've now changed one sentence back to how I would actually say it, compressed a lot of the paragraphing, corrected the date, and added this slightly ridiculous Anthropic detour because it occurred to me while reading the draft.
Which may tell us more about whether the system is becoming me than any voice score will.