The New Moat Is Taste

Most marketing content was weak before AI, and slow production hid how thin the thinking underneath often was. Bad copy used to take weeks and now takes seconds, and it was always bad.

The speed works as an audit. Every brand that was leaning on production time to disguise a weak point of view now has nowhere to hide. The agencies that were charging forty hours to produce a newsletter that said nothing are quietly finding that a solo operator with a real point of view and a good AI stack can produce a better newsletter in two hours.

The flood of bad AI content is real, and it is only part of the story. The critics are seeing the noise and missing the signal.

The old filter in content was production capacity: who could afford the writers, the designers, the project managers. That filter is gone. Anyone with a subscription can produce finished-looking output.

The new moat is taste: the ability to recognize a weak first draft, name what is missing, cut, and hold an audience clearly enough in your head that you can tell the model who it is writing for.

You develop taste through years of practice.

The moat is taste, and the skill is teaching the system to operate with yours. That takes four disciplines running in parallel, and none of them is particularly technical.

Real skill building

You have to do the work enough times to develop the judgment that teaching requires. If you cannot name a weak first draft when you see one, you cannot teach a model to avoid producing them. If you cannot articulate why a sentence is flat, you cannot correct it. Editing is a craft that pre-dates AI by a few thousand years, and no tool replaces the reps.

The operators teaching their systems well are the ones who have been writing, editing, and shipping for a long time, or who borrow from people who have. You are building the muscle that makes prompts useful.

Agent design

Stop treating AI as a conversation partner. Start treating it as a system you are designing. A chat starts over every time; an agent has structure. A chat needs you to re-explain your voice every session; an agent holds it. A chat produces one-off output; an agent produces consistent output at scale.

Give each repeat task its own small agent. One for research synthesis. One for social post drafts. One for client emails. One for proposals. Each with its own rules, its own reference materials, its own anti-patterns. The work is mostly organizing your practice.

If you find yourself re-explaining your voice to a model every Monday morning, you have built a habit of starting over.

Treat it like a smart junior with no context

The mental model that works is this: your agent is a capable junior writer who has read everything on the internet and nothing specific to you. They have skills. They have range. They have no taste of their own, because they have been trained on everyone's taste at once, which averages out to none.

Your job is to be the editor who gives them context. That means the audience, the voice, the job this piece has to do, and what is never going to fly. A good editor does this for a new hire in their first month. They coach, they reject, they point at examples, and they explain why a sentence is wrong.

Teaching a model works the same way: editing with intent, over and over, until the defaults move. The people who are bad at this have usually never been good editors. The people who are good at this have. It is a craft skill transferring into a new medium.

Prompt engineering around good, bad, and you

Most people prompt for what they want. The people producing sharp work also prompt for what they do not want, and most of all for what sounds like them. Three reference files, kept current and fed into every relevant agent:

• What is good: Sentences, paragraphs, and openers from my own work that I am proud of. The model learns my cadence and my rhythm from these.
• What is bad: Sentences I would cut. Corporate filler. Thought-leadership clichés. AI default patterns that have slipped through. The model learns what I am trying to avoid.
• What is you, or your brand: Specific formulations I use. Words I reach for. Structures I default to. Positions I hold. Phrases I would never use. If you are a brand, this is your point of view.

The third file is the hardest to write. Most people cannot describe their own voice. That is a separate problem, and solving it is a prerequisite to teaching a model anything useful. If your brand cannot say clearly what it sounds like, your AI system will default to industry average. Industry average is exactly what the critics are complaining about.

A few of my actual rules

Abstraction is the enemy here, so I will name some. These live in files I feed into every drafting agent I run.

• No em dashes. Ever: Em dashes were my favorite piece of punctuation for years. Now they are a tell. Period. Comma. Restructured sentence. I mourned them briefly. Now, nothing but irritation when I see them.
• No runway phrases: No "in today's rapidly evolving landscape" or any of its cousins. No "it is worth noting that." No "what this means is." Start with the idea.
• No corporate verbs: No "leverage," "unlock," "empower," "drive value," or "utilize." Use "use." Name the actual thing you mean. If the sentence collapses when you remove the filler, it had no content.
• No pre-announced enthusiasm: No "I am so excited to," no "couldn't be more proud," no "thrilled to announce." Confident people do not pre-announce their enthusiasm. They just say the thing.
• No abstractions for people: No "stakeholders" when I mean "marketers" or "founders" or "clinicians." Name the actual people.
• Vary sentence length: Short sentences get varied with longer ones. Five six-word sentences in a row read like a telegram.
• Tether claims to mechanisms: Every forward-looking claim gets tethered to a mechanism. "AI changes strategy" alone is not allowed. "AI changes strategy because the unit economics of delivery reset" is.

That is a sample. The full list is longer. Without a list like this, you are hoping the model teaches itself.

Most people approach this as a one-time task: write the rules file, upload it, done. It only works as a loop.

Every piece of work you ship is a new data point. Good pieces go into the reference file. Bad patterns that slipped through go into the anti-reference file. Every rule that stopped a problem once eventually starts blocking something useful, so rules are living and they need pruning. Every new project is a chance to notice a pattern you had not named yet. Name it. Add it.

The system you have in month twelve is a more articulate version of the one you set up in month one, pruned and extended from actual use. A rules file that has not changed in three months means you set it and forgot it.

I will say this one directly to the brand people reading.

Most brand books are useless to an AI. They are full of mood words, photography guidelines, and logo clear space. None of that teaches a system to write in your voice. The rules file for a brand holds the specific positions your brand holds that nobody else in your category holds, the sentences only you would write, the phrases you would never let out the door, the named frameworks that are yours, the anti-patterns in your category that you refuse to repeat.

If your brand does not have a file like that, your AI content is going to sound like your competitor's AI content. Both will sound like the category average. Write the file. It will do more work than any other document in your marketing stack.

A taught system stops producing generic work. It stops opening with "in today's rapidly evolving landscape." It stops describing things as "robust," "seamless," "cutting-edge." It stops writing sentences that could appear in any company's marketing. It starts producing work that sounds like you, because you have shown it what you sound like and what you do not.

I have built a small, specific editorial system around my voice, and I run my raw material through it until it comes out sharp.

Teaching like this rarely gets a keynote, and it shows up in the output.

The critique of AI content is usually aesthetic: it is ugly, it is soulless, it is generic. Those critiques can be true about a particular output and still miss the larger movement. Some critique the very concept, and they have a point: art is still art, and marketing copy is a different thing. The best of human writing can still surprise and inspire in ways AI cannot, at least for now.

The people doing the best work right now are editing past the complaints. They use AI to generate raw material. They apply the one thing AI still does not have, which is a point of view. They ship the result at a cadence that would have been unthinkable two years ago. And for the love of god, they read it after it is created. Really read it.

Taste plus iteration plus a systems mindset is the new operating stack. All three take practice.

The loudest complaints about AI content in 2026 will read the same way the loudest complaints about blogs read in 2006. True in detail, beside the point in direction.

Everyone can publish now. Almost none of it is good. The people who can use the speed, teach the system, and insist on a point of view are producing work that is faster than human pace and sharper than most of what agencies were shipping two years ago.

The field belongs to the operator behind the content.

The game is on.