Why are AI roleplay replies so short? How to get longer ones
Updated
If a reply ends cleanly but says too little, the model chose to stop; it was not cut off. It copies the length it sees: your turns, the greeting, the example lines and its own recent replies. Write fuller turns with something to react to, lengthen the card's examples, and leave room with a reply limit around 700 tokens. On UnsavedAI, Continue adds to a reply that stopped too soon.
Try it now, free, no account neededStart a story with Wade HarlanEight hundred miles of dust. One driver who won't talk.First, rule out a cut-off
Two different problems get called "short replies". If the text stops mid-word or mid-sentence, the reply hit a length limit, and why replies cut off has the settings. If it ends on a full stop after two lines, the model decided it was done. The rest of this page is about the second kind.
| What you see | What it means | Where to look |
|---|---|---|
| Stops mid-sentence | It hit the length limit | Reply length setting |
| Ends cleanly after a line or two | The model chose to stop | Your turns, greeting, example lines |
| Shrank over a long chat | It is copying recent thin replies | Reroll or extend a few of them |
| Short with one character only | The card is terse | That card's greeting and examples |
Your turns set the length
Models match the rhythm in front of them. Three one-line turns from you and the fourth reply shrinks to fit. With Wade Harlan, a driver who barely talks:
Your turn: "how far to the next town" Reply: "Two days."
Now give him something to react to: *I trace the cracked map on the dashboard. Four towns are crossed out in red.* "Did you cross these out, or did someone else?" *I tap the nearest one.* "Because Dry Fork is where you said we'd get water."
Reply: *Wade's eyes stay on the road for a long mile. Then one hand leaves the wheel and moves your finger off the map.* "I crossed them out." *Dust ticks against the windshield.* "Dry Fork's got water. It's also got people who remember this truck."
He is still a man of few words. The reply is longer because there was more to show: a gesture, the road, a reason to worry.
Lengthen what the model copies
The greeting is the model's first and strongest sample of a reply, so a 40-word greeting teaches 40-word replies. If you made the character, give it a greeting of 120 to 200 words that mixes action, setting and speech, and make at least one example line a full paragraph. In SillyTavern and JanitorAI those are the first message and example dialogue fields; in the UnsavedAI builder, Their first message and Example lines.
Inside a running chat, recent replies matter more than the card. If the last five are thin, reroll the thinnest or use Continue so the chat in view holds longer samples. SillyTavern has a Continue option too.
Settings give room, not orders
Max Response Length in SillyTavern and Max new tokens in JanitorAI are ceilings. Raising one from 300 to 700 lets a long reply finish, but never makes a short one longer. Temperature barely touches length. What does work is a line in the instructions the model reads every turn, the Main Prompt in SillyTavern or a proxy's custom prompt in JanitorAI: "Write two to four paragraphs per reply, mixing action, description and dialogue."
In any app, including UnsavedAI, one note does the same for a while: (OOC: longer replies from here, please: three or four paragraphs, with more of what you notice and do, not only what you say.)
When replies are too long instead
Reverse every fix above. Write shorter turns for a while, trim the longest recent replies, cut a greeting that runs past a screen, and lower the reply ceiling to 300 or 400 tokens if you like it punchy, accepting that the odd reply will need Continue to finish. Long replies also fill the context window faster, so early scenes drop out of view sooner. The note for this direction: (OOC: shorter now, one paragraph and a line of dialogue at most.)
Questions people ask
Why did replies get shorter the longer I played?
Each short reply makes the next one shorter. Reroll or extend two or three of the recent thin ones and the length climbs back.
Does a bigger model write longer replies?
Not by itself. Length follows the chat more than model size; a larger model mostly writes richer replies at a similar length.
My character is supposed to be quiet. Will longer replies ruin that?
No. Ask for more action and setting rather than more speech. A silent character can fill a paragraph with what they do.
