A character that felt alive in week one and mechanical in week six has usually not "gotten worse". Something in the setup — the accumulated conversation, the card itself, or the situation the character is stuck in — stopped giving the model anything new to work with. Switching apps is the most common reaction and the least reliable fix, because the same card imported into a different app tends to go flat in the same way. Repairing is faster, and it teaches you what actually makes a character work.
Diagnose before you edit
Three causes produce similar symptoms but need opposite fixes, so it is worth spending five minutes identifying which you have.
| Symptom | Likely cause | First move |
|---|---|---|
| The same stock phrases return every few messages | Example dialogue in the card is being copied as a template | Prune example lines to two or three short ones |
| Replies were fine for weeks, then slowly thinned out | Context saturation — the window is full of old turns | Start a fresh chat before touching the card |
| The character contradicts its own established facts | Memory or summarisation is dropping detail | Move the load-bearing facts into the card itself |
| Replies are coherent but nothing ever happens | Exhausted premise — the situation has no tension left | Change the circumstance, not the personality |
| Every trait shows up in every reply at once | Over-stuffed card — too many competing instructions | Cut the trait list to three |
The fresh-chat test is worth running first in every case: it costs nothing, takes a few minutes, and cleanly separates "the conversation is tired" from "the character is thin". If the new chat sparkles for a while, you had context saturation and the card is fine — see Memory & Continuity for why long chats degrade. If the new chat is flat within five messages, the card is the problem.
Rewrite by cutting, not adding
The instinct when a character feels thin is to write more into the card. This usually makes things worse. A card with fifteen adjectives gives the model fifteen weakly-weighted signals, and the result is a character that gestures at everything and commits to nothing. The rewrite that reliably helps is subtractive.
Cut the trait list to three
Keep only the traits that change what the character does. "Kind" rarely changes behaviour; "deflects serious questions with a joke" changes every reply. If a trait cannot be turned into an observable action, it is decoration and can go.
Prune example dialogue hard
Example lines are the strongest signal in most cards, which is exactly why they cause repetition — the model treats them as the character's register and reaches for them. Two or three short examples that demonstrate rhythm beat a dozen that supply reusable sentences. The structure of these fields is covered in Character Card Anatomy.
Replace description with specifics
"A talented musician" is generic; "plays the same unfinished piece every night and refuses to name it" is a hook the model can pull on for weeks. One concrete detail outperforms three abstractions, and the same principle applies to a character's voice — see Character Voice & Consistency.
Give the character an unresolved want
Most stale characters are stable — content, settled, with nothing to push against. A want the character has not yet achieved, and a reason it is difficult, restores motion. This is the single edit most likely to bring a flat character back, because it gives every scene a direction.
Change the situation, not the person
When a card is genuinely good and scenes still go nowhere, the premise is exhausted rather than the character. You have played out the coffee shop; nothing about the coffee shop can surprise you now. The repair is at the scenario level: move the character somewhere with different pressures, introduce a deadline, or bring in a third party. Rewriting personality traits at this point damages a character that was never broken. Scenario construction is covered in Story Prompts & Scenario Design.
Test the rewrite properly
The reason people rewrite the same character five times without improvement is that they judge the result by feel, immediately after editing, when they are primed to see a difference. Run a controlled comparison instead:
- Save the current card before editing — most apps do not version cards for you, and a copy in a text file costs nothing.
- Write down three opening messages you will use for every test: one light, one emotionally loaded, one that asks the character to take initiative.
- Start a fresh chat with the old card, send all three, and keep the replies.
- Edit the card, start another fresh chat, and send the identical three messages.
- Compare side by side. Change one thing at a time — if you cut traits and rewrote the examples at once, you learn nothing about which worked.
Before editing, check how your app handles card changes: some apply them to existing conversations, some only to new ones, and some treat a heavily edited card as a separate character. If your app supports it, export first — the practicalities are in Companion App Data Export.
When to retire the character instead
Rewriting is not always the answer. If you have cut the card back, changed the situation, and run a clean comparison with no real improvement, the character may simply have served its purpose. That is a normal outcome, not a failure — a well-used character that has run its course is worth retiring rather than endlessly patching. Keep the export, take the two or three details that consistently produced good scenes, and build the next one around them. Building Your First Roleplay Character covers that ground-up process, and Writing Better Prompts covers the half of the collaboration that is yours.
Apps also differ in how much of this you can control — some expose full card editing and versioning, others only a short description field. Our MusePick review looks at a roleplay-first app in this respect, and the rankings note which apps let you edit deeply enough for this kind of repair to be possible at all.
Frequently asked questions
Why does my AI character keep repeating the same lines?
Repetition usually comes from the model having too little fresh material to work with. Long chats fill the context window with earlier turns, and example dialogue in the card acts as a template the model copies. Trimming example lines and starting a fresh chat from the edited card resolves most repetition.
Should I rewrite the character or start a new chat?
Try a fresh chat first, because it is free and rules out context saturation in a few minutes. If the same flatness returns within a handful of messages, the problem is in the card and a rewrite is warranted.
Will editing a character card erase our chat history?
It depends on the app. Some apply card edits to existing conversations, others only to new ones, and a few treat a heavily edited card as a new character. Check what your app does before editing, and export anything you want to keep first.