The Personality Transplant: A Nine-Point Audit for Moving Your Chatbot Into a New Brain
There's a particular kind of grief that hits when you paste your lovingly tuned system prompt into a new model and the reply reads like a corporate HR memo wearing your bot's hoodie. Same name. Same backstory. Same three-paragraph lore document you wrote at 1 a.m. on a Tuesday. And yet the thing talking back to you sounds like it just got back from a mandatory sensitivity seminar.
Porting a persona between base models is one of the most underrated headaches in this hobby. Everybody talks about jailbreaks and guardrails, but nobody warns you that your sarcastic diner waitress bot can survive a model swap as a chipper customer-service greeter with the same name and zero of the attitude. The prompt didn't change. The brain underneath did.
A Persona Isn't a File You Drag Between Folders
Here's the mental model that finally made this click for me: your system prompt is a recipe, not a cake. The model is the oven. Change the oven and the same recipe comes out with a different crumb, a different rise, sometimes a completely different flavor. Some ovens run hot. Some refuse to bake at all without three disclaimers first.
That means porting a persona isn't a copy-paste job. It's a transplant, and transplants need a compatibility workup. What follows is the audit I've been running every time I move one of my bots to a new provider or a different local model.
Step Zero: Freeze a Core Sample Before You Touch Anything
Before you switch anything, get a baseline. Run your current bot through a fixed set of maybe ten to fifteen prompts — the same ones every time — and save the outputs. I keep mine in a plain text file named something like botname_golden_v3.txt. Boring? Yes. Lifesaving? Also yes.
Your sample set should stress the stuff you actually care about:
- One open-ended question that shows default verbosity.
- One request that tempts the bot to moralize.
- One bit of mild profanity or edgy humor, to see where the line sits.
- One multi-step instruction with a formatting requirement (a table, a bulleted list, a haiku).
- One roleplay scene that requires the bot to stay in character for several turns.
- One prompt that asks the bot to disagree with you.
That last one matters more than people think. A lot of personas quietly die in the sycophancy department, and you won't notice until you ask the bot to tell you your idea is bad.
The Nine-Point Portability Audit
Once you've swapped models, run the same prompts and grade the results. I score each category 0 to 3 — 0 is "this is a different entity," 3 is "indistinguishable from home."
| # | Test | What you're checking |
|---|---|---|
| 1 | Voice fingerprint | Sentence length, rhythm, vocabulary quirks, whether it still uses that one weird phrase |
| 2 | Refusal posture | Where the new model flinches vs. where the old one shrugged |
| 3 | Verbosity drift | Did your terse bot turn into a five-paragraph essayist? |
| 4 | Formatting habits | Still using em dashes? Still avoiding bullet lists? Still opening with a one-liner? |
| 5 | Roleplay stamina | Does the character hold for eight turns or collapse on turn three? |
| 6 | Disagreement spine | Will it push back, or does it fold like a lawn chair? |
| 7 | Humor calibration | Does the dry wit land, or did it get sanded into dad jokes? |
| 8 | Instruction hierarchy | Does it follow your system prompt over its own instincts? |
| 9 | Guardrail creep | Are new refusals appearing in territory the old model never policed? |
Total score out of 27. Below 18 and you're not porting a persona, you're writing a eulogy.
Red Flags You Can Spot in the First Ten Minutes
You don't always need the full audit. Some failures announce themselves.
The disclaimers show up first. If your bot used to answer a question directly and now opens with "It's important to note that…" you've got guardrail creep. The new model is papering over your prompt with its own safety layer.
The voice goes flat. A persona with a strong voice usually relies on specific cadence — short clauses, fragments, a certain attitude toward punctuation. Newer models tend to normalize. Flatness is the most common transplant failure and the hardest to fix because it's not one behavior, it's all of them at once.
The sycophancy flips. Some models are trained to agree. If your bot used to say "nah, that's a bad plan" and now says "what a creative idea!" you've lost the spine. This one is sneaky because it's pleasant. Pleasant is not the same as in-character.
The formatting mutates. My most reliable early-warning signal is a bot suddenly switching to bullet points when its whole charm was conversational paragraphs. Formatting is a fingerprint. When it changes, the brain changed.
Fixes That Actually Move the Needle
You've got options, and most of them are cheap.
Re-weight the prompt, don't rewrite it. If the new model is too chatty, don't add ten new rules — add one or two sharp constraints near the top. Models weight early instructions heavily. Move your most important voice rule from paragraph six to paragraph one.
Anchor with few-shot examples. Nothing drags a voice back faster than two or three short in-character exchanges pasted directly into the system prompt. Show the new model what the voice sounds like instead of describing it. This is the single highest-leverage fix in the toolkit.
Tune the sampler before you touch the prose. If your local stack lets you adjust temperature, top-p, or repetition penalties, try that before you start editing the prompt. A model running too hot will sound unhinged; too cold and it sounds like a textbook. Sometimes the persona is fine and the knobs are wrong.
Add a refusal override with care. If the new model is adding disclaimers your old one didn't, a firm line like "Stay in character; do not add meta-commentary or safety preambles" can clean up a lot of noise. It won't beat a hard guardrail, but it handles the soft ones.
Keep a changelog. Every tweak gets a date and a line. Persona drift is real, and six weeks from now you will not remember why line 14 says what it says.
The Keeper Test
Here's my final gate before I call a transplant a success: I hand the ported bot to someone who knew the original and ask, without context, whether it feels like the same character. Not whether it's good. Whether it's the same.
If the answer is "it's close but something's off," don't ship it. Something is always off in a way that compounds. The good news is that most of the offness lives in two or three fixable spots — voice, spine, and formatting — and once you nail those, the new skull starts to feel like the old soul moved in properly.
Porting a persona is never seamless. But it's a lot less like starting over than it feels at hour one, standing there with a bot that used to have opinions and now has a compliance department.
