← Back to census
Progress note rail · Guide
Talk with the Note

Generate the morning note from the chart dump, then work the note in words: tell it what to change, leave to see the patient, come back and pick up the same conversation. One thread per patient per day, kept on the server as you go.

1
Where it is

Generate a progress note on the rail as usual. On the review screen, under the note, is the Talk with the note pane: the conversation so far on top and a box to type in at the bottom. Drag its top edge to give the note or the conversation more room; Hide folds the thread away and leaves the box.

When you come back to the patient later in the day, the rail's input screen offers Today's note with a Pick it up button: the note as the conversation last left it (or as last saved), with the whole thread, and no second generation.

It is there for progress notes only. It is not on the H&P or discharge summary (the Admissions workspace has its own version for the H&P).

2
Say the change

Ask mode and Talk to the chart were retired on 10/1: talk is for changing the note, not for questions.

Type one instruction and press Enter (Shift+Enter for a new line). Plain words work. Examples that worked on real turns:

creatinine is 2 today, stop fluids
changes the labs line, the AKI problem, and the septic shock plan that said continue fluids
lovenox 40 daily, not heparin
one line in DVT prophylaxis, nothing else
she's full code, put that in the disposition
add hyponatremia, sodium 129, fluid restrict to 1.2 liters
adds a new problem block at the end of the plan
update the retrosternal chest pain, i gave the whole history in my update
naming the problem asks for it: the block is rewritten from your update and the chart

Aim first when you mean one block. Press Aim on a note card (beside Refine and Edit, which work as before). The card lights, the box says Aimed at with its name, and your next change touches only that block. Select a sentence in the card before you press Aim and the change is about that sentence. The ✕ on the chip clears the aim; with no aim, the note works out which blocks your words concern, as before. The aimed problem also lights in the Course column on the left, and clicking a problem name there brings its note card into view.

A change takes about 4 to 25 seconds. The button reads Changing while it works. You can keep reading the note; if you edit it while a change is running, your typing wins and that change is not put in. Each turn sees the turns before it, so "and the same for the potassium" works.

3
Read the trail

Each thing you said appears in the thread as a card: your words, then what happened. When you leave the box, the newest card stays in view above it.

creatinine is 2 today, stop fluids ✓ CHANGED · AKI · GROUNDED Cr 2.0 today. Stop IV fluids. LEFT ALONE: 'Cefepime day 4.' in Septic shock, not asked, matches chart keep it

Lit words are what this turn changed. Right after the turn, the change is also lit inside the note cards above until they redraw, and a line the turn took out shows under its card as Removed (on screen only; it is not note text). Hovering an edited sentence there shows the one it replaced.

GROUNDED means the changed lines were checked against your words and the pasted chart.

The problems named after CHANGED are the blocks that changed. If a problem you meant is not listed, say it by name ("the septic shock plan too").

A change made with an aim shows → the block's name beside your words. When your words also concern other blocks, the card names them and says they were left alone. Aim at one of them, or clear the aim, to change it.

Lines the turn held back are tagged. LEFT ALONE means the model changed or added a line you did not ask about, so the note keeps your earlier version (or leaves the new line out); tap keep it to take the model’s version after all. LEFT OUT means the line was not note text; only a cleaned part, if any, went in, and there is no keep it. The label says why: not asked (or not asked, matches chart) when the model changed a sentence in a problem you did not name; points to another note, placeholder or filler when the line was not note text ("see speech pathology notes", "[plan needed]", "Monitor closely"). Your own words are never left out. When anything was left out, the reply says what the note actually holds.

not in chart is a chip on a change you did ask for whose claim is found nowhere in your words or the pasted chart ("baseline 0.9" when no baseline is charted). It stays in the note; read it before you sign.

Not checked. means the checker was unavailable for that turn. The note is exactly what the model returned; read the changed problems yourself.

4
Undo, versions, save

Undo next to the box puts the whole last change back. Every change also adds a Told chip in the Versions row, so you can flip between before and after like Tighten.

Every turn is kept the moment it is answered, with the note text it produced, under this patient and today's date. Close the rail, regenerate, switch patients and come back: the thread is still there. When you press Save Draft or Sign, the turns also ride along with that note. Nothing is written to the chart, to Note Evolution, or to another patient. Tomorrow starts a new thread.

5
What it will not do

The hospital course is never rewritten by this. Ask for it and you get a warning; edit the course by hand.

It does not write placeholders, pointers to another team's note, or plan lines you did not give; when something is not known, the problem leaves it out.

Generate works the same way: anything the model wants to tell you about the draft shows in the ◆ About this draft box above the problem cards, never inside the note. A line marked TAKEN OUT was removed from the draft because it talked to the reader ("Please review…", "per provider"); if its fact belongs in the note, type it in yourself. Copy and Save Draft leave the box out.

It does not invent values. If you say "add the troponins", they come from the pasted chart. If you state something the chart does not have, it is written as you said it, because it is your note.

It does not know what you meant if you did not say it. "Fix the plan" changes nothing useful; "stop the lisinopril while the creatinine is up" does.

What is going on underneath. Your instruction is first read by a fast classifier that answers one yes/no per block of the note: does this instruction change what this block should say? Only the blocks that come back yes are handed to the model, so it cannot touch the others. The model returns new text for those blocks. Then every changed sentence gets two more yes/no answers: was this asked for, and is every claim in it found in your words or the chart. A sentence that was not asked for goes back to your prior text and shows as Left alone. The whole check adds under half a second. With an aim the block is your choice: the classifier does not choose, it only names the other blocks your words touch.
If something looks wrong: tap Undo, or pick the Original chip in the Versions row. Nothing you said is lost; the thread is kept for the day. Tell the developer what you typed and what the thread said.