Building Taste
Into the Machine.

I run a weekly newsletter.
Here is what I built to keep it alive.

Most people use AI to do things faster.
That is the wrong ambition.

The right ambition is to preserve what only you can provide. Speed is a commodity. Taste is not.

Three hours of Slack.
Forty open tabs.
One tired opinion.

That was the old process: scan channels, chase links, resolve paywalls, make editorial calls at midnight. The newsletter shipped. The voice did not always survive the journey. I did not want to do it faster. I wanted to stop losing the thread of what I actually thought.

What Caught
My Eye.

Ten links, every Friday, in my voice. I ran this newsletter for years on Revue. Twitter bought it and shut it down. This is the rebuild: same editorial instinct, tools I own this time.

I am not a developer. Claude built the infrastructure. I directed the editorial logic.

Score the candidates
before you see them.

Fifteen candidates ranked by originality, voice fit, publisher quality. The algorithm's top ten are auto-ticked. I only override what I disagree with. The default state already reflects judgement; I am editing, not selecting from scratch.

After eight issues,
it knows what I like.

The scoring algorithm reads eight weeks of pick history. Domains I consistently pick get a bonus. Ones I consistently ignore get a penalty. No retraining. No prompt engineering. Just a feedback loop built into the score.

Wednesday night I write the takes.
Thursday morning they arrive as the copy.

The picker pre-fills a take field for each candidate in my editorial voice. I edit or keep it on my phone. Thursday the build agent reads those takes as the foundation for the full newsletter copy. AI sharpens; it does not invent. Ten weeks in, the copy arrives sounding like me before I have touched it.

Here is what
it looks like.

Fifteen candidate cards, each scored and ranked. Category colour bars across the top. A take field pre-filled in my voice below each headline. Swipe to pick on mobile; tap to check on desktop. The default top ten already ticked. I am editing the algorithm's answer, not starting from scratch.

What it actually
runs on.

Not a product. A workflow. Each tool doing one job.

The system carries the weight.
Taste carries the voice.

Taste is the moat. AI does not replace editorial judgement; it preserves it at scale, if you build the right layer between the algorithm and the output.

The specific stack will not transfer. The pattern will: score the candidates, learn from the choices, preserve the voice.

A strategy consultant scoring fifty research links a week would run the same layer: score by client relevance, weight by which angles actually made it into decks, pre-fill the brief framing. A pitch writer. A buyer at a culture magazine. Anywhere a human is making the same kind of choice repeatedly, the loop can learn it.

The bottleneck is no longer technical skill.
It is knowing what you want clearly enough to ask for it.

The layer
gets deeper.

Ten weeks of picks. Each one adjusts the weights. The system now reaches for what I favour before I ask. That is memory: not storage, but preference that compounds.

The next loop adds framing. Thursday's copy trains Wednesday's take suggestions. The system learns not just what I pick but how I write about it. Two layers of compounding, not one.

Not faster.
Sharper, every week.

Yes.
And the taste compounds.
Ten issues. Voice sharper now than when I started.
Not because I improved.
Because the system remembered.