OOC is not one disease. It is three, and they need different medicine

A knight-commander who has spoken in clipped, formal sentences for two hundred messages suddenly says “I hear you, and your feelings are valid.” The player screenshots it, posts it with one word — “OOC” — and gets twelve replies recommending twelve different fixes. Rewrite the card. Get a better model. Lower the temperature. Start over.

Most of that advice is wasted, because “out of character” is a lump diagnosis. In the support cases and experiments we have run, it covers three separate failures with three separate causes, and medicine for the wrong one does nothing. We wrote about whether AI roleplay writing holds up at all in an earlier reply to a reader; this piece is the methodology that was too long for that letter, generalized from fanfiction to any long-running roleplay.

Triage first. One question per cause: Did the character ever sound right? If no — from the very first messages — it is cause one, the card. Did they break at a specific point, contradicting something established earlier? Cause two, the window. Did they erode gradually, fine last month and generic now? Cause three, the loop. Timing is the diagnosis, and it takes thirty seconds of scrolling your own history to make it.

Cause one: “the card is detailed, but the AI ignores it”

The complaint is almost always literal: the card really is detailed. Two thousand words of backstory, childhood, motivations, a tragic betrayal in act two. And the character still opens with the warmth of a customer-service macro, because none of that prose told the model how this person talks.

Models follow demonstrations far more reliably than descriptions. The fix is structural, and the chara_card_v3 spec has fields for exactly this. Three to five example exchanges, chosen because they carry the character’s voice — those lock tone better than any adjective list. Hard rules for surface habits: forms of address, sentence length, the tic that appears when they are cornered. Then negative constraints, the most underrated line in any card: “never apologizes first”, “never explains her motives unprompted”. In our testing, one prohibition prevented more drift than ten positive traits, because a prohibition is checkable and an adjective is not.

Concretely, for the knight-commander: delete “she is stern but secretly caring, shaped by loss” and replace it with one exchange where she answers a plea for comfort with logistics — “You will eat. Then you will sleep. Grief keeps until morning.” — plus two rules: addresses everyone by rank or surname; never asks about feelings, acts on them instead. Fifty words. That is more enforceable than the entire backstory essay, because next turn the model is looking at a working example of the register it must produce.

Dosage: cut before you add. A few hundred words of core persona plus demonstrated dialogue outperforms the essay version, and the essay actively hurts on cost — the whole card is re-sent every single turn. If your card format is fighting you on where any of this goes, the field guide to card formats maps the slots.

Cause two: “she suddenly acted like a stranger”

Different symptom, different failure. The character was fine for eighty messages, then greeted a sworn enemy like a friend. Players read this as personality collapse. It is usually amnesia: a model reads a fixed budget of text per reply, your card plus recent history, and the vow of vengeance from message twelve stopped fitting long ago. The model did not defy the established dynamic. It never received it this turn.

The card survives every turn — it is re-sent in full. Loose history does not. So durable facts need the same durability: a lorebook, keyed entries injected only when their keywords surface in recent chat. The vow, the debt, who knows whose secret. Dosage rules from the three tickets we debugged: telegraphic entries around 150 words, facts first, no atmosphere — an 800-word literary entry starved everything else out of the injection budget in one real case. Keep constant entries under five, because each one is rent paid on every message. And put every alias in the keyword list; in our support-case tally, missing aliases explained more than half of all “never triggers” reports. A card keyed to a full name nobody uses in chat is a card that never fires.

This slice of the problem borders on the wider memory question — windows, summaries, who holds the transcript — which we covered separately in the piece on why roleplay chats forget. Here the point is narrower: mid-arc personality flips are usually supply failures, not character failures.

Cause three: “a month in, everyone sounds the same”

The slowest one, and the one no card edit fixes. Week one is brilliant. Week four, the sardonic mercenary, the shy scholar, and the feral witch all speak in the same smooth, even register — the model’s own.

The mechanism is a feedback loop. Every reply the model writes goes back into the context and becomes a demonstration for the next reply. The more of the history is model output, the more the model imitates itself instead of the character. We measured this in 20-round continuation experiments: the original author’s sentence-length variation scored 0.84, and all six models sank to between 0.46 and 0.59 as their own text filled the window. Different task, same loop — a roleplay chat is that experiment running at conversational speed.

The same experiments carry a hopeful result: the loop responds to intervention. Labeling which passages were canon and which were model output, with one carefully worded instruction, took one model from 20-out-of-20 failure rounds to zero. You cannot rewrite our prompt internals from a chat window, but you hold the equivalent levers. Curate ruthlessly: a drifted reply you keep is a lesson you just taught. Swipe it, regenerate it, or edit it into voice before moving on. Re-anchor occasionally — write a message of your own in the story’s register, or bring back a line the character said in week one; recent demonstrations are exactly what the model leans on hardest. And when a loop has already set in, switch models mid-chat: a different model does not inherit the previous one’s self-imitation groove, which is why we treat model switching as a repair tool, not a luxury.

Which fix goes first?

The three causes are independent, but the treatment order is not. Card first, one evening: example dialogue, speech rules, prohibitions. Without it, you cannot even tell the other two failures apart, because nothing ever sounded right to begin with. Lorebook second, an hour: the ten facts that must never be contradicted, keyed to every alias in actual use. Curation is not a step, it is a habit that starts on day one — cheaper than any repair, because a clean history never needs one. And model choice comes last, not because it is worthless but because it is the only lever that costs money to experiment with and the only one that fixes nothing on its own.

What none of this fixes

The verdict. Whether “I hear you, and your feelings are valid” is out of character for your knight-commander is a judgment only someone who knows the character can make — the tools put candidates side by side, inject the right facts at the right time, and keep the loop from eating the voice. They do not know who she is. You do. That division of labor is the whole design: structure what can be structured, and keep the taste decisions human.

One admission to end on: cause three has no set-and-forget cure. Every model we tested drifts toward its own register given enough of its own output to imitate. Curation is maintenance, not a one-time fix — the cost of a long roleplay that still sounds like itself in month three is a swipe here and there in month one.