Short version: in Gemini, Claude, and Grok, Enter sends the message. You want a new line, and the assistant already has half a sentence and is busy answering it. I built myself a chat window that works the other way round. Enter makes a new line, I lay the instruction out in points, typos get fixed as I type, and I can see the limit before I send anything.
How annoying is it? VERY!!! You write a five-point request, your finger does what it does in every text editor, and… gone. The model politely answers the first line. You delete, add the rest, and pay twice for the same question. Out of a limit that always runs out at the worst possible moment.
The window is called Mostek, which is Polish for a ship’s bridge, so from here on I’ll call it Bridge. It is an upgraded interface for an AI assistant that runs on my own computer (OpenClaw) and talks to Claude, Grok, Gemini, or local models. It is not a new model and not an app you can buy. It is the same chat window, fixed in the places where an ordinary chat window gets in the way. For now it runs only on my machine.
One caveat up front: this is not a model review. Sonnet, Gemini, and Grok will swallow a mess and build something out of it. The trouble starts earlier, in the little box where the mess gets typed.

🎬 The video and the short
⏎ Enter is not a command
In an ordinary chat, Enter means “send”. A new line lives under Shift+Enter, and in a terminal it depends on the program and on the terminal itself. Your head knows that. Your hand does not, and it presses Enter.
In Bridge, Enter stays in the text. Sending takes three Enters in quick succession, up to 650 ms apart. The blank lines from that burst disappear, so the model does not get a tail of empty rows. Ctrl+Enter sends at once, on a phone or when you are in a hurry. Holding Enter down sends nothing.
No Bridge? Do at least one thing: write longer requests in a notes app and paste them in finished. Shift+Enter in the browser works too, as long as you remember it on every single line. I don’t.
📋 Points you can actually read
A dash at the start of a line makes a bullet, a number makes numbering. a) b) c) walks the alphabet. After a number, a dash or a letter, Enter adds the next point instead of sending, and an empty point ends the list. A slash on an empty line opens a ready prompt structure: Goal, Context, Constraints, Result. Each section is a heading you can fold.

This is what it looks like when I actually type. Carelessly, without the accents, like a monkey at a keyboard. In a numbered list, Enter does not fire the command: it sees a point and makes the next one. Correction runs as I go, so the field ends up with text I can read. The grammar mistakes in my notes are on purpose. The typos are not: I notoriously eat letters and swap them around.

Why bother, if the model will understand a wall of text anyway? So that you understand your own request tomorrow morning. An instruction is information, and information you cannot read is information you cannot check. You do not know what you actually asked for, so you cannot tell whether the result is any good. I wrote about information as material rather than decoration in Everything Is Information.
The field grows with the prompt, up to about 60% of the window height, so I see the whole prompt at once. A long instruction stops being a two-line strip where the third point vanishes past the edge.

All of these tweaks are for me, not for the model. They help me put my thoughts in order before they become a prompt for the AI assistant. I no longer write long prompts in a separate text editor and paste them into the chat. I write them right in the chat window.
✏️ Proofreading for you, not for the model
The model will swallow a typo. You, rereading your own request, should not have to. After a full stop, a question mark, or Enter, the last sentence goes off to be corrected. Grammar and syntax get fixed; irony, tone, and word choice stay. I do not write with a smirk so that a proofreader can turn it into an official memo.
A small change flashes and goes in, and Ctrl+Z takes it back. A big rewrite that changes the meaning never goes in by itself: you get a suggestion and click it if you want it. Code, addresses, paths, mentions, and commands are left alone.
The correction is not another message in the conversation. By default Grok does it, and when Grok is unavailable, a local model on my computer. I can switch the engine with one click. Gemini 3.5 Flash-Lite is on the list too: it barely touches my API budget, and it is fast and good. There is also LanguageTool, which fixes typos only, with no language model at all. The Claude limit I am chatting on stays untouched.
🎛️ A model, or “not now”, without leaving the window
The chip above the field shows the session model. Auto means the default model plus backups: when Claude hits its limit, the conversation moves to another model and carries on. A pinned model has no backup, and switching models mid-conversation throws away the prompt cache, so the next message costs more. The window tells you before you confirm, not after.
At the end of the text, arrow down and three Enters open the choice right there: a model for one message, or “where to”. Chat now, queue, local queue (no cloud, works offline too), heavy-job queue, an agent that edits files, or Grok. The “when” field takes a date and time. A queued job waits until the limit allows it, and the result arrives on Telegram.

Example: it is 11 p.m. and the limit window is nearly empty. You schedule the job for 7 a.m. and go to bed, instead of hovering over Enter waiting for the reset.
⏱️ The limit sits above the window, and sessions can be found
Above the chat are two bars: the 5-hour window and the week. The weekly figure comes from readings. Without a fresh reading the bar is an estimate, and it is labelled as one. I would rather see an estimate up front than get a surprise halfway through an instruction.
On the session list, every conversation has a context ring. Below 20% it is blue, then green, yellow, and orange, and from 80% it turns red. Red means: time to start a new conversation before the model starts losing the beginning. Hover over a session and you see the model, the context percentage, and a lower-bound estimate of what the next question will cost. Next to it is the energy, water, and CO₂ footprint, also as an estimate. Why I count energy and not just dollars is in Energy Currency.
The rule is simple: one session, one topic. I sort sessions into groups and search them by content, archived ones included. I do not have to remember which tab from a month ago held that one sentence.
🪟 The ordinary window stays
Bridge locks nothing in. The “Original” button switches the tab back to the stock panel, with the same session groups and names. If you prefer an Enter that fires off half a sentence, nobody is taking it away.
What is not there yet: automatic model choice based on the content. Today Auto is simply the default model with backups, no guessing.
Polish version: Okno czatu AI: Enter zostawia nową linię.
❓ Short answers
Does Enter send the message in Bridge? No, it makes a new line. Sending is three quick Enters or Ctrl+Enter.
Does autocorrect use up the conversation limit? No. A separate engine does it: Grok by default, a local model, Gemini 3.5 Flash-Lite, or LanguageTool. A small change goes in by itself, a big one waits for a click.
Where can I see how much limit is left? Above the window: a 5-hour bar and a weekly bar. On the session list, a context ring, and a cost estimate on hover.
Can I install Bridge? Not yet. Would you like to? Tell me in the comments.