You ask AI which subject line is better. It answers with total confidence. You use its pick, and the email reads like everyone else's.
To teach AI your taste, stop asking it which option is better. Tell it which one you picked, and why, in one line. Keep those reasons as a dated list of rules it reads before every job. Then let your customers decide which rules stay.
That list is a Taste Log. It's how to make ChatGPT less generic, or Claude, or Gemini, without writing a style guide.
Below: the five steps, the free prompt I tested on my own data, and where to save the log.
Why AI Keeps Handing You the Safe Pick
AI hands you the safe pick because that's what it's built to predict. A model trained to be broadly helpful "reaches for the safest answer" when you leave out who it's for, as this breakdown of why ChatGPT gives generic answers puts it.
So I tested it on my own list.
I've run 12 subject-line A/B tests in beehiiv between August and October 2026. Two were ties. I gave a fresh Claude chat the subject lines, with no results and no tools, and asked it to guess which line my readers opened more.
Leave out the ties, and Claude matched my readers 8 times out of 10.
The catch: only one of those tests was big enough to mean anything: 86 people on each side. Claude got that one wrong.
Claude picked the line that tells the whole story. My readers picked one sharp moment with a number on it. AI picks what usually works. Your customers aren't usual.
Fair to Claude: my list is small, and it flagged sample size itself. It told me to check how big each test group was before trusting its picks, and warned that Apple Mail privacy can make open rates look higher than they really are.
Still, the AI can't know your readers until you tell it what they chose.
What Is Taste in AI?
Taste, defined: Taste is what you pick and why: which opening, which price, which name, which client gets a yes. AI can't learn it from your writing samples, because samples only show what you kept. They never show what you turned down, or the reason.
The second half matters most. When Lon Riesberg wrote up teaching an AI his taste, he found that "the rejects, and especially the near misses, turned out to matter more than the picks."
Taste vs Voice: The One-Line Difference
Voice is how you sound. Taste is what you pick.
AI gets close to your voice from a few good samples. I covered that side in how many writing samples it takes to train ChatGPT on your voice.
Taste is what samples miss. A perfect copy of your voice can still choose the wrong subject line, in your exact tone. That's the whole taste vs voice gap.
It's why "make it sound like me" never fully fixes generic output. You get the draft, then you edit it to make it sound human in parts. The choices underneath are still the average ones.
How to Teach AI Your Taste: Five Steps
Five moves: Spread, Pick, Say why, Judge, Log. The AI spreads and keeps the list. You pick. Your customers judge.
Spread: Ask for Options from Obvious to Rare
Do this: Ask for several options, ordered from the most obvious to the rarest, and ask how likely each one is.
There's research behind the "how likely" part. In the Verbalized Sampling paper, asking for several answers with their probabilities gave 1.6 to 2.1 times more variety in creative writing than a plain ask (Zhang et al., arXiv 2510.01171).
You should now have: a short list where option one is the line every rival would send.
Check: if the rare end looks like the obvious end reworded, ask again for options that make different promises.
Seeing the safe one is the point. Now you can skip it on purpose, or keep it on purpose.
Pick and Say Why in One Line
Do this: Choose one. Then write one line on why, in your own words.
Not "it feels better." Write something a stranger could apply next time, like "owners skim, so promise something they can use." Adele Newenham gives the same advice for teaching Claude your taste: next to every example, write the pattern behind the label.
You should now have: a reason. That reason becomes your rule.
Check: would your reason choose between these exact options if you weren't in the room? If both options fit it, it's too loose.
Let Your Customers Judge, Not the AI
Do this: Use your pick. Send the email, post the offer, make the call. Then wait for the result.
Don't ask the AI whether your pick was good. Felix Rieseberg of Anthropic said it plainly in his October 2026 essay on working with models: "Do *not* ask the model to judge itself."
The AI knows what usually works. Only your customers know what works for you: opens, replies, bookings, a yes.
You should now have: a result next to your pick, not an opinion.
Log It so AI Starts from Your Picks
Do this: Keep one rule per line. Newest at the bottom, each with the date and the result once you have it.
A log line looks like this:
[date] · Pick [your rule] · result: [opens, replies, yes or no]Paste the log at the top of your next request, or save it in your AI's settings (see below).
You should now have: an AI that starts from your rules instead of the average.
The Free Prompt That Teaches AI Your Taste
The prompt below runs the whole loop in one paste. It's #233 in my library: Taste Log: Teach AI the Choices You Would Make. I tested it in Claude on my own subject-line data, and it passed.
Swap the [bracketed bits] for one real choice you've made.
You help me write down my taste, so AI stops handing me the safe, average option.
Here is a choice I made:
What I was choosing: [a subject line, a price, a product name, a proposal opening...]
The options I had:
[option 1]
[option 2]
[more, if you had them]
The one I picked: [your pick]
Why, in one line, in my own words: [your reason]
What happened, if I know yet: [opens, replies, a yes or no from the client, or "not yet"]
My Taste Log so far: [paste it here, or write "none yet"]
Now do five things:
1. Tell me which option most people in my shoes would pick by default, and why, in one line. That is the safe choice. Say if my pick is the same one.
2. Turn my reason into one rule I can reuse next time. Start it with "Pick" or "Choose". Make it about my business and my customers, not general advice.
3. If a rule already in my Taste Log points the other way, quote it and ask me which one wins. Do not decide for me.
4. Give me my updated Taste Log: one rule per line, newest at the bottom, each with today's date and the result if I gave one.
5. If the result went against my pick, say so in one line. Then suggest one rule the winning option points to, marked "suggested", so I can keep it or drop it.
Use only what I told you. Do not add rules I did not give you, apart from the one marked "suggested" in step 5, and do not tell me whether my pick was good. The result decides that.What the steps are there for:
Step 1 names the safe choice. On my big test, it named the $47 line. That's the same line Claude picked blind, and the one fewer of my readers opened. If your pick matches the safe one, ask yourself: is that me, or habit?
Steps 2 to 4 do the bookkeeping. Your reason becomes a rule, clashes with old rules get flagged for you to settle, and the log comes back dated.
Step 5 is for when your customers disagree with you. On my big test, it suggested this rule from the winning line: "Choose the subject line that shows one specific moment of the AI-bill pain, like five charges in one day, and saves the full story for the email." It's marked "suggested", so you decide whether it goes in.
Save Your Taste in ChatGPT, Claude and Gemini
The prompt works in any AI that takes text: ChatGPT, Claude, Gemini, DeepSeek. I tested it in Claude.
Save the log where your AI reads it before each chat. If you've searched how to teach ChatGPT my preferences, this is the answer: your preferences are the log.
ChatGPT: Custom instructions, or a Project's instructions. Custom instructions apply to every chat. A Project keeps related chats, files and instructions together. (overview)
Claude: Instructions for Claude, or Project instructions. Project instructions only apply to chats inside that project. (Claude help)
Gemini: Instructions for Gemini, or a Gem for one recurring job. Google says Gems move to skills on personal accounts from November 2026. (instructions, Gems)
Menu names as of October 2026. They change, so keep the master copy in a document you control.
How to Make AI Less Generic for Good
AI hands you the safe pick because the safe pick is what it predicts. Better voice samples won't fix that, because taste lives in what you choose and why.
A blind Claude matched my readers 8 times in 10 and still missed the one test big enough to count. Closing that gap is how your work stops getting lost in a sea of LLM-generated slop.
So run the Taste Log prompt on one recent choice, then save the log it gives you.
The free Taste Starter adds a prompt for options from obvious to rare, a check on whether your AI knows your taste yet, and a blank Taste Log. The Full Kit is $15, and the rest of my tested prompts live in the Vault.
Frequently Asked Questions
Why Does ChatGPT Give Generic Answers?
ChatGPT gives generic answers because it predicts the most likely reply, and the most likely reply is the most common one. Models have seen many of the same websites, so with no audience, constraints or past picks to go on, they fall back on safe defaults. A Taste Log gives it your picks to start from instead.
Can AI Learn Your Taste?
AI can learn your taste inside what you give it, but it doesn't learn you permanently. It reads the rules you saved, each time. Keep the log yourself and it travels to any tool.
Does This Work with Claude and Gemini Too?
Yes. The Taste Log is plain text, so it works in Claude, Gemini, ChatGPT, DeepSeek or any AI that takes text input. I tested the prompt in Claude.
How Many Choices Before It Works?
There's no set number. One choice gives you one rule you can use on the very next job. The log gets more useful each time a real result lands on it, and the near misses teach it most.
