Persona, System Prompt, Fine-Tuning: The Three Layers
An AI’s personality is built in three layers: a persona document written for the character, a system prompt written by the operator, and fine-tuning applied to the model itself. Two of those are text, resent with every message. The third changed the model’s weights in a training run that finished before you ever opened the app. How far any edit reaches depends on which layer it lands in.
Key Takeaways:
- A persona document and a system prompt are the same mechanism, text placed in front of the conversation. What separates them is who wrote it, what order it sits in, and whether you are allowed to read it
- Nothing in the architecture marks instruction text as more important than the rest of the input, since the design relates any position in the sequence to any other (Vaswani et al., 2017)
- A system prompt outranks your persona because models are trained to weight that slot heavily, not because the software enforces it
- Fine-tuning changes the model’s weights in a training run owned by whoever ships the model. It lands on every user at once and no consumer setting reaches it
- A behavior that survives an emptied persona in a conversation with no history lives in the weights, and rewording will not remove it
How AI Personality Is Built, Layer by Layer
Two of the three layers are files and one is the model itself, which is the whole reason the three behave so differently under editing. A persona document and a system prompt get read fresh on every request and can be rewritten in a minute. Fine-tuning is not read at all. It is baked into the numbers doing the reading.
| Layer | What it physically is | Who writes it | Where it sits in a request |
|---|---|---|---|
| Fine-tuning | Weights inside the model, adjusted by a training run on example conversations | Whoever trains and ships the model | Nowhere. It is the thing reading the request |
| System prompt | Instruction text the app puts ahead of every conversation on the platform | The company operating the app | First, before anything you wrote |
| Persona document | Character text: name, manner, history, sample lines, rules | You, or the app’s writers | After the system prompt, ahead of the messages |
No save button starts a training run. That is a scheduled engineering job with a hardware budget behind it, run against the model long before it reached any app, and it is why “training her” is one of the most misleading phrases in this category. Filling in a character form is writing. Talking for 5 months is filling a database. Not one of those actions moves a single weight, which is why the model answering you on day 150 is the identical model that answered on day 1. It stays identical until the operator swaps in a new one, which is what a model upgrade does to a character you know: the text is untouched and the thing reading it is not.
Which gives you a rough test for any claim a product page makes. If a feature would require changing the model, and the company is not in the business of training models, the feature is text.
Why a System Prompt and a Persona Are the Same Trick
Both are strings pasted in front of your conversation, and the model has no way of knowing one came from a company and the other from a text box you filled in on a Sunday. It receives a single sequence. The design behind these models relates every position in that sequence to every other one when it produces each piece of a reply (Vaswani et al., 2017), so an instruction at the top and a sentence you typed 40 messages ago are the same kind of material to it.
That has a consequence people find hard to accept. There is no protected channel. Nothing at the level of the machinery says this part is rules and that part is chat, which means a system prompt does not have authority in any structural sense. It has position, and it has habit.
The habit is where fine-tuning comes back in. Models are tuned on examples where the instruction at the top gets followed and the user’s later contradictions do not override it, so the model learns to treat that slot as binding. Layer 2 works because layer 3 taught it to work. Strip out the training that installed instruction-following and a system prompt becomes decoration.
Two practical things fall out of that. Your persona sits below the operator’s text in the same stack, competing on the same terms, which is why an app can quietly override anything you wrote by editing a file you never see. That same operator-controlled layer is where added behaviors live, including the scheduling that makes an AI text you first. And the instruction stays a fixed size while the conversation grows around it, so its share of the input shrinks with every exchange.
Before you decide a character is badly written, check whether she is merely outnumbered.
What Only Fine-Tuning Can Change
Everything that shows up before you write a single word of persona is in the weights. Default reply length. The habit of closing with a question. The particular way it hedges, the register it falls back to when the brief is thin, how literally it takes an instruction, whether it obeys a rule at message 3 and forgets it at message 30. The pull toward agreeing with whatever position you have already stated was tuned in the same place (Sharma et al., 2023), which is why no line in a persona reliably removes it.
Instruction-following is itself one of these trained behaviors, and that produces the ceiling nobody warns you about. How well it obeys your persona is a property of the model. Past a certain point, a sharper instruction stops buying anything, because the limit was never how clearly you expressed the rule.
Here is the part people get backwards. They assume vague behavior means a vague brief, so they write more: longer backstory, harder rules, capitals, three restatements of the same instruction. That makes the fixed block bigger without making it stronger, and it crowds out the conversation the model needs to answer what you just said. More text is the most common wrong answer in this whole category.
So when a fix does not take, stop editing and find out which layer you are fighting. Rewriting layer 1 to defeat layer 3 is a losing evening.
Why People Expect a Toggle to Do a Training Job
Because product copy calls all three layers “personality,” and two of the three respond instantly. You move a slider, the next reply is warmer, and the lesson you take away is that this settings page controls what the character is. Then you hit something that will not move at all, and there is no vocabulary on the screen to explain why.
Nothing on a settings page reaches the weights. A trait control is a text edit with a nicer interface: some value becomes a word or a sentence in the same block of instructions you could have written yourself. Whether the character reads as a romantic partner or a platonic AI friend is the same kind of text edit, changeable in a sentence. Against a trained tendency, it is a suggestion arriving in the same envelope as everything else.
The damage from the mix-up is not just wasted effort. People conclude the app is broken, or worse, that they described her badly, and both readings send them back to the text box for another hour. A behavior that comes from the model would have survived every one of those rewrites.
Ordering Your Edits So You Stop Guessing
Work from the cheapest test to the most expensive, and 10 minutes settles most of it.
Start with dilution. Open a conversation with no history, paste the identical persona, send the same 3 messages that went wrong. If the character comes back sharp, you were watching your own accumulated tone win, and the answer is shorter chats rather than a bigger brief.
If the behavior survives that, empty the persona down to 2 lines and send the same 3 messages again. Whatever is still there is not yours. It belongs to the operator’s text or to the model.
Now separate those two. Create a second character in the same app, as different from the first as the form allows, and send the same messages a third time. A quirk that shows up on both characters is coming from the layer they share, and you have just found the floor of what that app can be. Worth noticing at this point: switching apps changes layer 2 and usually not layer 3. Unrelated companies rent the same handful of models.
What does not work is asking her. An AI companion will produce a confident account of its own configuration on request, and that account is composed the same way every other reply is composed. It is writing, not a readout.
FAQ
What is the difference between a system prompt and a persona? Both are instruction text sent with every message; the system prompt belongs to the app and sits first, while the persona belongs to the character and sits after it. Priority runs top down, so the operator’s rules win where the two conflict. Most apps hide the system prompt entirely, which is why a character can suddenly refuse something she did fine with last week.
Can you fine-tune an AI companion yourself? No consumer companion app offers it, because fine-tuning means running a training job against model weights, and the resulting model would apply to everyone using it. What creation tools call training is writing a persona and letting a memory store fill up. Both are text handed to an unchanged model.
Why do two characters I made sound the same? Because they share the two layers you did not write. The same system prompt sits in front of both, and the same model generates both, so whatever the pair has in common is the app’s floor rather than a failure of your descriptions. Making the two briefs more different will not move it.
Does the persona still work after the app changes models? The text is untouched, but the thing reading it is not the same, so the identical brief can produce a noticeably different character. A description that leaned on one model’s habits, like a dry manner it fell into naturally, has to be written out explicitly for the next one. Sample lines survive model changes better than adjectives.
The three layers are ranked in an order almost nobody notices: the one you can edit freely has the least authority, and the one with the most authority is the one no user, and often nobody at the app, can touch. A settings page inverts that ranking, because it can only show you the parts that are yours. Once you can name which layer a behavior came from, half the frustration goes away and the other half turns into a real answer, which is sometimes that this app cannot do what you are asking and no amount of writing will change it.
Lona runs on the same three layers as everything in this category: two of them text you can edit, one a model it did not train. Next time the character feels off, that split is the honest place to start, because it tells you which layer you are actually fighting.
Sources
- Vaswani, A. et al., “Attention Is All You Need,” 2017
- Sharma, M. et al., “Towards Understanding Sycophancy in Language Models,” 2023
