Picture a follow-up visit for hypertension. The patient says the new pill is making them cough. A scribe that only hears the conversation writes "patient reports cough on new medication". A clinician writes "dry cough since starting lisinopril 10 mg on August 14; ACE inhibitor cough likely; switch to losartan". The difference is not intelligence. It is context: which pill, since when, at what dose, and what else is on the list.
That is the problem Terra's Context Engine solves. Before the draft is written, Terra gathers the chart information you allow, labels it clearly, and shows it to you. This post explains what goes in, what stays out, and the rules that keep yesterday's chart from becoming today's findings.
Today first, the chart second
The order matters. A draft in Terra always starts from today's evidence: the context you write for this visit, the reviewed transcript, the template fields you fill in, and any documents you select. At least one of these must exist. Chart context alone never starts a draft, because a note written only from the chart is not a note about this visit.
Then the chart sources are added, each as its own labelled source with its own citation ID.
The encounter header
Every draft gets a short structured header: the patient's age on the date of service, pronouns, visit type, date, clinician and note language. Age is calculated from the date of birth on the date of service, so a patient seen the week of their birthday gets the right number.
What the header does not contain is just as important: never the patient's name, never their MRN, never their date of birth. The model does not need them to write a good note, so it does not get them.
Allergies and active medications
Approved allergies and active medications are always included when they exist. Not every allergy someone has typed somewhere: only entries in the patient's Facts that a clinician has reviewed and approved, and whose source is still current. If a fact came from an outside document and that document's text has since been corrected, the fact is no longer current until someone reviews it again.
These two lists are always on because getting them wrong is costly, and because the draft regularly needs them: to name the medication the patient is describing, to avoid suggesting language that conflicts with a documented allergy, and to write a medication section that matches the chart.
Active problems, when you want them
Active problems are included only when you turn them on in the visit's source picker. For a focused visit about a sprained ankle, the problem list is noise. For a chronic-care visit, it is the backbone of the assessment.
Prior notes, when you choose them
You can include the last one to three verified signed notes for the patient, newest first, each with its amendments. Only signed notes count: a draft from another visit that nobody has signed is not evidence of anything.
Historical, not today
Here is the rule that took us the longest to get right. Every chart source except the encounter header and today's facts is labelled, in the prompt and in the citation, as historical. Prior notes carry the label "Historical — prior encounter, not today's findings". The problem list carries "Historical — chart problem list, not today's findings".
The instructions to the model are explicit: historical sources are never today's history, examination, results, assessment or plan. If prior information is relevant, it must be attributed, for example "Per prior note dated 2026-08-14, blood pressure was 148/92", and it must be cited.
Why so strict? Because the most dangerous error an AI scribe can make is a plausible one. "Lungs clear to auscultation" copied forward from last month's note reads exactly like a finding from today. It is not one. By labelling history as history, and requiring attribution, Terra makes copied-forward findings visible instead of invisible.
Two related rules:
- Recorded is not the same as true. Chart facts are reported as recorded entries. The draft says "penicillin allergy recorded (rash)", not "patient is allergic to penicillin", unless today's conversation confirms it.
- Absence is not a finding. If there is no allergy entry, the draft never writes "no known allergies". An empty list means nobody recorded one, not that the patient has none. NKDA is something a clinician confirms, not something software infers.
Shown before it is sent
When you choose Prepare draft, Terra shows a quote: the estimated credits and every source the draft will use, by name and kind. Allergies and medications, problems if enabled, each prior note by date, each selected document. If something is listed that you did not expect, change the source picker before you confirm. Nothing reaches the model without appearing in that list first.
When the chart changes mid-draft
Drafts take a little time. Charts keep moving. A nurse approves a new allergy while the draft is running. A colleague amends last week's note. A typo in the date of birth is fixed.
Terra pins the exact version and hash of every chart source when the draft starts. Before the result is applied, it rebuilds the context and compares. If anything changed, the draft is stale. Terra charges the usage the provider already spent, discards the draft and leaves your saved note exactly as it was. You prepare it again against the current chart.
That can feel strict when the change was small. We think it is the right default. A draft that contradicts an allergy approved thirty seconds ago is worse than no draft.
The record brief is separate
Patient profiles also have a record brief: literal excerpts of reviewed facts, approved documents and signed notes, for you to read before the visit. The brief is for you, not for the model. It is never pasted into the draft, so there is exactly one, inspectable path for chart information to reach a note: the labelled sources in the quote.
What this adds up to
Context is what turns a transcript into a note a clinician would actually sign. But context is also where an AI scribe can quietly go wrong. Terra's approach is to include what matters, label what is historical, show everything before it is sent, and throw away a draft rather than apply it to a chart that has moved on.
If you want to try it, start a trial with a synthetic patient, approve an allergy and a medication, and prepare a draft with and without the prior notes. The Help Center has the step-by-step version.
