Why the assessment and plan reads like a summary of the visit

Before you touch the prompt, read the transcript for one of the offending consultations and look for the plan. Not the note — the raw transcript. In most clinics that do this exercise honestly, the specifics are not there. The clinician said the tablets he is on are fine for now, we will see how you get on, come back if it gets worse. The model turned that into a plan that reads as a recap because that is a faithful account of what was said. Nothing has gone wrong with the generation; the evidence for a plan does not exist in the recording.

This is the one documentation complaint that a better prompt makes worse rather than better. Push a model harder for specificity when the specifics are absent, and it fills the gap — with a plausible dose, a plausible interval, a follow-up nobody agreed to. The section that fails against missing evidence by hedging is behaving as designed. Fix the evidence.

Audit 10 consultations against 5 commitments

1. Pick 10 recent encounters from one clinician on one list. Mixing specialties or clinicians at this stage tells you nothing, because the speaking habit is the variable you are measuring.
2. For each, open the transcript and the signed note side by side. You are comparing what was available against what was produced, which is a different exercise from grading the note.
3. Mark each of the 5 commitments as spoken, implied or absent. Spoken means a named drug at a named dose, a stated interval, a stated review point, stated safety-netting advice, a stated onward referral.
4. Count implied and absent separately. Implied — 'carry on as you are', 'the usual dose' — is the interesting column, because it is where a model has to choose between hedging and guessing.
5. Compare the counts across the 10. If the same commitment is absent in 8 of 10, you have a capture problem specific to that element, not a general quality problem.

The output of that hour is a table you can act on. It is also the only version of this problem anyone can argue about productively, because 'the plans are vague' is a feeling and 'the review interval was never spoken in 8 of 10 consultations' is a finding.

The 5 elements, and how often each is actually said out loud

That last row is the one teams underestimate. A referral decision is frequently made in the 40 seconds after the door closes, when the clinician is weighing a waiting list against a clinical threshold. No microphone in the room will ever hear it, and no prompt will recover it. It has to be captured somewhere else or accepted as absent.

4 reasons the commitments never reached the recording

• The decision was made silently. Experienced clinicians resolve a lot without narrating it, particularly on continuations. There is nothing to fix in the clinician; there is something to change in the workflow.
• The decision was made after the patient left. Referrals, imaging requests and dose adjustments that depend on looking something up happen at the keyboard, outside the recorded window.
• The clinician writes the plan by habit rather than speaking it. For anyone who has typed notes for 15 years, the plan is a writing act, not a speaking act, and ambient capture does not change that reflex on its own.
• The template gives the plan no required fields. A free-text plan box asks for prose and gets prose. A section with named slots asks for commitments and exposes the empty ones.

Note what is not on that list: mishearing. If the dose was said and the note has it wrong, that is a transcription or terminology fault with a different diagnosis and a different fix, and it usually shows up in drug names across several clinicians rather than in one section. Similarly, if whole stretches of the consultation are absent from the transcript, the plan being thin is a symptom rather than the disease — where consult audio actually dies is the page to run first, because a plan spoken in the last 3 minutes of a recording that stopped at minute 9 is not a plan problem at all.

3 fixes, and which one your audit points at

1. A spoken close-out. The clinician says the plan to the patient before they stand up: what you are taking, for how long, when I want to see you, what to watch for. 20 to 30 seconds, and it produces evidence for 3 or 4 of the 5 commitments at once. Choose this when your audit shows the decisions were made in the room but not voiced.
2. Structured capture at signature. A short form at review — review interval, referral destination, follow-up owner — that writes into the plan section as discrete values rather than prose. Choose this when your audit shows the decisions are genuinely made after the patient leaves.
3. A template with required slots. The plan section cannot be marked complete while a slot is empty; empty slots print as an explicit 'not discussed' rather than as silence. Choose this when the note structure itself is inviting a paragraph instead of a set of commitments.

A plan section is not a writing problem. It is a record of decisions that were either spoken aloud or were not, and no amount of instruction will make a model hear something the room never said.

Designing the plan section so an empty slot is visible

The template change is small and worth doing regardless of which fix your audit points at. Replace the single free-text plan box with named slots — medication, investigation, referral, review, safety-netting — and give each a defined empty state. An empty slot must render as an explicit marker, not vanish, because a section that quietly shrinks to fit its evidence is indistinguishable from a section where nothing was decided. The same structural principle stops a generated note filling a section it was given nothing for, which is the mechanism behind normals appearing in the review of systems that nobody asked about.

If the clinic already runs a paper proforma with a plan block on it, that block is your starting structure and it encodes years of local practice. Converting it properly — keeping the clinical intent, dropping the fields that exist only because paper cannot branch — is set out in turning a paper proforma into a template a model can fill. Do that before writing a new template from scratch.

Rolling out a speaking change without losing the clinic

A close-out ritual is a behaviour change, and behaviour changes fail in the same predictable way: they are introduced across a whole department at once, by email, with a rationale about data quality. Introduce it with 1 or 2 clinicians who want the notes to be better, run the 10-consultation audit again 2 weeks later, and show them their own before-and-after. Which clinic to start with is a real decision with a real trade-off, argued in the busiest clinic or the willing one.

Where the fix is structured capture at signature rather than a speaking change, it is a small piece of workflow software: a form that appears at review, writes discrete values, and is faster to complete than typing the same sentence. That is ordinary internal tools and ops work, and it lives or dies on whether it adds seconds or removes them. Everything around how a draft becomes a signed note sits in ambient documentation and the clinical note, inside our healthcare and MedTech work.

• assessment and plan
• ambient documentation
• clinical templates
• capture design