AI repeats my message back to me? How to stop the echo
Updated
The echo, where a reply restates your message before answering it ("So the bridge is out..."), comes from assistant training: models learn to confirm what they heard. In a story it wastes the start of every reply and slows the scene. Trim the restated part from recent replies, write turns with one clear thing to react to, and add a line asking for reactions, not recaps. On UnsavedAI, Edit trims it in seconds.
Try it now, free, no account neededStart a story with Dr. Maya EllisonNight shift. Full waiting room. You're her new resident.What the echo looks like
You write: *I drop the soaked map on the table.* "The bridge is out. We go through the pass or we wait a week."
The echoing reply: *She looks at the soaked map you dropped on the table. The bridge is out, and the two of you either take the pass or wait a week.* "So the bridge is out. The pass, or a week." Only then, if at all, does anything new happen.
This is different from the AI repeating its own phrases. An echo copies your words, usually in the first lines of the reply, every turn. When it retells your actions in detail, it edges toward speaking for your character as well.
Why models do it
Assistant training rewards showing that the request was understood, so a model's first instinct is to restate it. Roleplay writing adds a second habit, a reply that "picks up" the partner's last action before going on. And the more there is to restate, the more it restates: a long turn packed with several actions and questions invites a point-by-point answer, like a reply to an email.
Before and after
With Dr. Maya Ellison on a night shift where you are her new resident:
You: "Bed four says the pain started at dinner and it's worse when he breathes in." Echo: *Dr. Ellison listens as you report that the patient in bed four has had pain since dinner, worse on breathing in.* "So: bed four, pain since dinner, worse when he breathes in. I see." No echo: *Dr. Ellison is already walking.* "Worse on breathing in. Did you listen to his chest, or are you guessing?" *She pulls back the curtain on bed four.* "Show me."
The second reply uses three of your words and spends the rest on her: a challenge, a decision, a scene that moves.
Fixes in any app
- Edit the restated opening off the last few replies. The real reply usually starts at the second or third sentence.
- End your turn on one thing for them to answer, not three questions. Less to summarise, less summary.
- Put your character's action last, so the reply has to begin by responding to it.
- Paste once: (OOC: don't restate what I wrote. React to it from your character's side and move the scene forward.)
There is one place where echoing belongs. Negotiators deliberately repeat a person's last few words to keep them talking, and Lieutenant Dana Kessler, on the phone with you while you hide in a server room, has every reason to. In character, on purpose and a few words long, it works. A paragraph of recap every turn is the habit to remove.
Where to put a lasting rule
In SillyTavern, a line in post-history instructions works well because it sits right after the newest messages: "Do not summarise or repeat {{user}}'s last message; continue from it." In JanitorAI, a proxy's custom prompt is the place for the same line. In a card you wrote, check the example lines: if each {{char}} reply begins by repeating the {{user}} line above it, the echo is being taught. Rewrite them to open on the character's own move. On UnsavedAI, those live in Example lines in the character builder; on house characters, trimming the first two or three echoing replies usually ends it.
Questions people ask
Why does the echo get worse when I write long messages?
There is more to confirm, so the model spends the start of its reply confirming it. Shorter turns with a single focus get more direct answers.
Can repetition penalty fix it?
Only a little. It discourages reusing your exact words, so the echo turns into a paraphrase. Editing the habit out of the chat works better.
