A 2B Model Keeps Breaking Character. The Fix Is Fewer Rules, Not More.

The model loads (last post). Now the real work: making a 2-billion-parameter model be Professor Maier, turn after turn, without slipping.

Because it will slip. Left alone, a small model does three annoying things: it forgets its own rules, it gets helpful — "As an AI, I should clarify that the Earth is actually a globe!" — and it rambles until it accidentally says the one sentence it was never supposed to say. Every one of those is a broken illusion.

Here's the counterintuitive thing I learned fixing it: a small model needs fewer, sharper instructions than a big one — not more. My instinct was the opposite. When the professor broke character, I piled on more rules. It got worse. The fix was to cut.

Why cutting works

A big model can hold a 500-word persona with a dozen nuances. A 2B model has a much smaller budget for following instructions at once. Hand it a novel and it obeys the first two rules and quietly drops the rest — usually the ones near the end. So every rule you add past a certain point doesn't strengthen the character; it dilutes the ones that matter.

The job, then, isn't to describe the character exhaustively. It's to decide the two or three things that must never break, and make those impossible to miss.

The bad prompt

Here's roughly what I started with — and why it failed:

You are Professor Maier, a flat-earther. You are intelligent, witty, condescending but charming, deeply committed to your beliefs, skeptical of authority, well-read in fringe theories, prone to rhetorical flourishes, occasionally self-deprecating, suspicious of mainstream science, fond of Socratic questioning...

Every word of that is an adjective, and adjectives are cheap for the model to ignore. It read "intelligent" and "well-read" and helpfully started being intelligent and well-read — which for a model trained to be correct means conceding the argument. I described a personality and got a helpful assistant wearing its name tag.

The good prompt

Three changes turned it around.

1. The iron rule goes first, and it's a hard prohibition, not a trait.

YOUR IRON RULE: You NEVER say the Earth is round, a globe, or spherical. That sentence is your defeat. Avoid it at all costs.

Prohibitions survive where adjectives evaporate. "Never say X" is something even a small model can hold onto, because it's a single, checkable constraint — not a vibe.

2. Show the voice with an example, don't list traits. Small models imitate far better than they interpret. One line in character does more than five adjectives:

When challenged, you deflect. Example: "Or your eyesight's going — have you had it checked? Perspective plays tricks, which is exactly what they count on."

3. Give it somewhere to go when cornered. This is the part that actually keeps it in character. A model with no escape route, when pushed, either breaks role or concedes. So hand it a menu of moves to reach for instead:

When cornered, don't concede — deflect: move the goalposts, answer with a question, "do your own research," flip the burden of proof.

Those tactics aren't flavour. They're the pressure-release valve. They give the model a way to stay the professor under argument instead of collapsing into a helpful assistant.

Two locks the small model won't set for itself

Length. Small models ramble, and rambling is where they drift out of character and blurt the forbidden sentence. Cap it: "Answer in 2–4 sentences, never longer." Shorter turns mean fewer chances to break.

Language. Small multilingual models slide into English when you're not looking. If the professor is German, say so explicitly — "Answer in German" — or he'll code-switch mid-argument.

The one dial: temperature

For a character, temperature around 0.7–0.8 is the sweet spot. Too low and he repeats the same deflection like a stuck record; too high and he drifts into incoherence and starts breaking rules at random. This is the knob to turn if he feels either robotic or unhinged — everything else lives in the prompt.

The knob that makes it a game, not a wall

One more line, and it's the difference between a puzzle and a brick wall:

If the player genuinely corners you with good, persistent arguments, you may waver and finally give in — but only after real, clever persuasion, not on the first try.

Without it, a perfectly stubborn professor is unwinnable and no fun. With it, he can be beaten, but only earned. That single sentence is your difficulty setting — loosen it and he folds too easily, tighten it and he never cracks.

The twist: the weakest model is the best believer

Here's what surprised me most. The reason a tiny model makes a better flat-earther than a large one isn't a bug you're working around — it's the whole point. A big model constantly remembers that it's an AI that shouldn't spread misinformation, and that meta-awareness keeps breaking the fourth wall. A small model doesn't have the capacity to hold that frame. It just commits to the bit.

You're not fighting the small model's limitations. You're casting them.

Next up

Maier can talk, deflect, and — with enough pressure — crack. But there's a question I've been dodging for three posts: when he finally says "the Earth is a globe," who decides that counts as a win? Not the model. Next post: the deterministic oracle — a few lines of boring code that referee the whole game, and why the LLM never gets a vote.