All Things AI · Season 3 · What You Can Build · Episode 1
You can train a model on writing you have already done until the draft comes back in your cadence. It is not a style prompt. It is a small adapter trained on your own work, it runs on the computer you own, and it costs nothing after the first afternoon.
The problem with asking a model to sound like you
Every agent has tried it. You paste three of your emails into a chat window and ask it to write the next one in the same voice. What comes back is close enough to recognise and wrong enough to rewrite, and after four rounds of that you write it yourself and decide the whole idea was oversold.
The reason is simple. A prompt is a description of your voice. The model reads the description, then generates from everything it has ever seen, and the average of everything it has ever seen is stronger than your three examples. You are asking it to imitate. It is going to approximate.
Training is a different mechanism entirely. Instead of describing your voice in the request, you adjust the model itself so that your patterns are what it reaches for by default. It is no longer imitating you from a description. It is producing from a version of itself that has been shaped by your work.
What an adapter actually is
The technique has a name you will see everywhere once you know it: LoRA, which stands for low-rank adaptation. The useful part is not the acronym.
The useful part is that you are not retraining a whole model, which would take a data centre. You are training a very small set of additional numbers that sit alongside the model and lean it in your direction. That small set is the adapter. It is a file, usually a few dozen megabytes, and it belongs to you.
It is small enough to run on your own machine
A modern laptop with enough memory can train one in a few hours and run it instantly afterwards. There is no cluster, no cloud account, and no per-use charge. This is the part that surprises people who assume model training is something only large companies do.
It is yours and it is portable
The adapter file is an asset. It does not live inside somebody else’s product, it does not stop working when a subscription lapses, and it can be pointed at a different base model later if a better one arrives. Almost nothing else in an agent’s software stack has that property.
It captures rhythm, not facts
This is the part most people get wrong. An adapter learns how you write: sentence length, where you put the emphasis, what you never say, how you open and how you close. It does not learn facts about your business, and it should not be trusted to. Facts come from your files. Voice comes from the adapter. Keeping those two jobs separate is what makes the whole thing reliable.
What you need to train one
Less than you would think, and you almost certainly already have it.
- Thirty to fifty pieces of your own writing. Scripts, newsletters, long social posts, client emails you are proud of. More is better, but this is the floor where the difference becomes obvious.
- The writing has to actually be yours. Anything a tool generated for you will teach the adapter to sound like that tool, which defeats the entire exercise.
- One consistent kind of writing beats a mixture. An adapter trained only on your video scripts will write scripts beautifully. One trained on scripts and emails and captions together will write all three adequately.
- A laptop with enough memory to hold the model. This is the only hardware question, and the episode on local models covers it.
How the training actually goes
You collect the writing into a folder. You pair each piece with a short description of what it was for, which is what teaches the adapter the relationship between a request and your response to it. Then you run the training, which is one command, and you wait.
The waiting is the longest part and it is unattended. A few hours on a laptop for a first adapter is normal. You are not sitting there.
When it finishes you have a file. You point the model at it and ask for something, and the first time you read the output there is a specific feeling of recognition that no amount of prompt engineering produces. That is the whole reason to do this.
The rule that makes it useful instead of frustrating
One pass, then correct one thing.
The failure mode is treating the output like a rough draft and rewriting the whole piece. If you do that, you have spent an afternoon training a model in order to keep writing everything yourself.
The correct workflow is to read what comes back, find the one beat that is wrong, and change only that. Not because the rest is perfect, but because the rest is already yours, and rewriting it wholesale is how you accidentally sand your own voice off. If more than one or two beats are wrong every time, the adapter needs more of your writing, not more of your editing.
It is also worth being clear about what the adapter is not. It carries your cadence, not your judgment and not your rules. It will happily write a sentence you would never say about a subject you never touch. Every check you would apply to your own draft still applies to this one.
What it grows
Consistency, which is the thing a referral business actually runs on.
Most agents are excellent in bursts. The first month of a new content habit is sharp and the sixth month is whatever got typed on a Sunday night. An adapter removes the correlation between how your week went and how your writing sounds, because the starting point is always your best register rather than your most tired one.
It also removes the tradeoff between volume and voice. Publishing more has always meant sounding less like yourself, because the shortcuts you take at volume are the shortcuts that flatten you. This is the first thing that genuinely breaks that tradeoff.
Where this goes wrong
Four failure modes, and every one of them is avoidable if you know about it in advance.
Training on text a tool wrote for you. This is the most common and the most quietly destructive, because the material looks like yours and reads fine. What the adapter learns is the tool’s cadence with a light coating of your topic choices, and you end up with a machine that reliably produces the exact thing you were trying to get away from. If you are not certain a piece is yours, leave it out.
Mixing every kind of writing you have ever done. An adapter trained on scripts, emails, captions, and long articles together learns the average of four different registers, and the average of four registers is a register you do not actually have. Train one adapter on one kind of writing. Train a second one later if you want a second register.
Expecting it to know things. It does not know your market, your business, or what happened last Tuesday, and when it is uncertain it will produce something plausible in your voice, which is considerably more dangerous than something implausible in a generic one. Facts come from your files every time.
Rewriting the output. If you rewrite the whole thing you have not saved any time, and worse, you have started editing your own voice back out of it. Change the one beat that is wrong. If several beats are wrong on every attempt, the answer is more training material, not more editing.
What it costs
Nothing per month and nothing per use. The training runs on hardware you already own, the adapter is a file on your disk, and running it afterwards is free. The cost is one afternoon of setup and the hours of unattended training, and that cost is paid once.
Questions agents actually type
How much of my own writing do I need to train an AI on my voice?
Thirty to fifty pieces is the practical floor where the difference becomes unmistakable, and one consistent kind of writing works far better than a mixture. It has to be writing you actually did, not writing a tool produced for you.
Is training a model on my writing the same as prompting it with examples?
No, and the difference is why prompting disappoints. Examples in a prompt are a description the model reads and then averages against everything else it knows. Training adjusts the model so your patterns are what it reaches for by default.
Does an adapter trained on my writing know facts about my business?
No, and it should not be trusted to. It learns rhythm and phrasing, not information. Facts come from your files and your memory layer. Keeping voice and facts as separate jobs is what makes the system reliable.
Why does my AI writing still not sound like me after I gave it examples?
Because examples in a prompt are a description, and the model averages that description against everything else it has learned. Training changes the model itself so your patterns are the default rather than a request. It is a different mechanism, and it is the reason prompting plateaus where training does not.
Season 1 is Claude Code for Real Estate Agents, filmed and on its own hub. All three seasons: AI for real estate agents.