🧂 Taste is automate-able

Most weeks, my own setup refuses to save a draft. It’s nothing exotic: a few plain text files my AI reads before it writes, and a small script that checks what comes back.

It flags the sentence, names the rule it breaks, and holds the file until the line changes. So I change it, and the version I land on sounds less like a machine and more like me.

I keep that friction deliberately, because it solves a problem I keep running into on client projects. A team has AI producing content at a volume no real-life editor can reasonably check, so the quality criteria start slipping. The drafts stay competent and on topic, and they could belong to any of four companies in the category.

The usual diagnosis is that the model isn’t good enough yet. The usual fix is better prompts. But I think it’s more about (yes, I’ll say it!) taste.

The part everyone assumes can’t be handed over

Taste is the word we reach for when we describe what’s missing, and it ends the conversation. Taste sounds innate. Something you either have or don’t, and certainly not something you can put in a file and hand to a machine.

Most of what I’d call taste is a pile of decisions I already made and never wrote down. Which ideas I’d defend in a room. Which ones are technically correct and still not mine. A model arrives with all of the marketing writing in the world. However, it holds no record of what I think about any of the hairy marketing problems.

Building an AI-assisted content system means writing it down and loading it into the tools you write with, in the three places where the judgment actually happens: picking the idea, drafting it, and deciding whether the draft is good enough.

Picking what’s worth writing

A model will produce ideas all day. What it can’t do is tell you which ones deserve your name, and if you ask, it waves nearly all of them through.

You already make that call. Watch yourself do it and try to reverse-engineer the questions that run through your mind at that step. Is this the version anyone in my category would write, or the version only I can? Have I actually watched this happen? Can I name the case, with a client and a number attached? And do I believe it, or does it just sound bold?

Those questions are the part you can codify. Mine sit in a file, and every idea goes through them before a word gets written. A few get killed there. More often they come back sharper, because the questions force the specific version out of the vague one.

The AI’s job at this stage is to ask you, on the days you’re rushing, the questions you’d put to yourself on a good day. It doesn’t need an opinion of its own.

Drafting, with your fingerprints on it

Two things belong to you at the drafting stage: what you never want to see, and what you actually sound like.

The first is the AI clichés blacklist, and it’s mostly not a vocabulary problem. That’s the common misunderstanding about these lists. People strip out every “delve” and “unlock”, then wonder why the writing still reads as if a machine produced it.

(That paragraph got this very edition rejected by my AI, by the way. My list doesn’t care that I was quoting the two words rather than using them, so before the file would save I had to sign off a written exemption for two verbs. I’m choosing to read that as the system working.)

The real tells are structural:

• Three short sentences in a row, for false gravity.
• The two-word question that exists to set up its own answer.
• The little aphorism parked at the end of a section, engineered to be quotable.
• The false reframe, where the second sentence exists to correct the first.

The second key part is your voice, described concretely enough to be followed. The sentence lengths you actually use. The jokes you’d make and the ones you’d never. Most voice guidelines stall here, because they offer adjectives, and no model can act on “warm, bold and human”. You need the examples from your real writing to bring it all to life.

What you get is a first draft that doesn’t sound like everybody else’s, which is a much shorter walk to where you want to end up. I’d be wary of anyone promising more than that in one pass.

The loop that asks you questions

When it comes to editing, improving a draft means improving it against something. With no standard attached, a model moves toward a general idea of good writing, averaged from everything it has read. Clearer, smoother, more like everything else.

Your own standard is more specific than you think. You know within two paragraphs whether a piece is working. You know which weakness you’re most likely to forgive at 5 pm on a Thursday.

Written down, that becomes a scoring rubric: the handful of things a piece of yours has to do, each one specific enough that a draft can fail on it. “Engaging” and “on-brand” can’t reject anything. They’re hopes. “Names a real case in the first third” can.

Then you let it run a couple of rounds against that rubric. What comes back is a better draft, plus a list of questions only you can answer, about the claims, the examples, the real customer situations you can reference. That’s where a piece stops being generic, because the answers come out of your projects and the things you’ve watched go wrong.

***

All of this makes my standard portable. The writing is still mine, start to finish. It respects my quality criteria not just when the AI is writing, but when I’m all out and rushing and would just want this thing DONE.

The obvious risk is a standard that rots. You can set a formal review cadence to fix that, though in practice you’ll keep adding to the criteria as you go, because you get too tired of correcting the same mistake every week.

Once you can see what each part is for, you can build your own version in whatever tool you already use. The questions come out of how you already judge ideas, the voice out of what you’ve already published, the rubric out of the corrections you keep making twice.

I’ve been pulling my own apart to teach it, ahead of the Content Stack Setup starting on 7 September, which is why it’s all so present for me right now and how that “codified taste” idea came to me in the first place.

The standard matters,

👌 Handpicked stories for you

Here are some interesting things I stumbled upon this week while reading, writing, researching, and being a generally curious marketing person on the internet:

💼 Three for work

How to Write Content That Lands With Decision Makers | Contently names the four decisions executives are actually making while they read your content: defending a budget, build versus buy, the risk of doing nothing, and telling vendors apart. Building content that solves them is your whole job.

Why Your Distribution Strategy Is Stunting Your Audience Growth | Robert Rose’s case for slicing content across channels instead of resizing one asset for all of them: the provocation goes on rented platforms, the proof on earned ones, the depth on what you own.

Google Doesn’t Punish AI Content; It Punishes Bad Content (331k Pages Studied) | Useful context for the AI watermarking noise of the past few weeks. According to the study, only 5.3% of top-three results are fully AI-generated (which means no-POV AI slop is rightfully invisible), but AI-written pages hold stable impressions instead of falling off a cliff (which means AI writing and human editing with a human perspective works).

👀 Two for fun

This prime example of artistic perspective | Monet and Renoir set up at the same riverside in 1869 and came away with completely different afternoons. One painted the water, the other painted the crowd.

This visual fix for messy recipes | I never knew recipe tables were a thing and I love how simple and concise the format is!

🐈 Weekly cat

When your internal mountain lion is let loose.