MindAudit — 4 April 2026 — Clinical Practice

AI Documentation Prompts for Psychiatry Registrars: What Actually Works

The challenge with AI in psychiatric documentation isn't getting the AI to write something. It's getting it to write something clinically structured, defensible, and actually useful when you're an hour past handover.

The documentation burden at the end of a psychiatric shift

Psychiatry registrars carry a documentation load that is qualitatively different from most other specialties. A single inpatient review might require a mental state examination, a risk formulation, a biopsychosocial summary, a medication rationale note, and a handover entry. In the emergency department, a thorough psychiatric assessment note can run to 600 words or more before you've started on the plan.

The cognitive weight of this isn't just the writing. It's context-switching: you've just spent an hour with someone in acute distress, and you now need to produce a structured, third-person clinical record that accurately represents what happened, reflects your clinical reasoning, and would withstand review by a consultant or legal examiner years later.

Most registrars develop their own templates over time, often inherited from terms, seniors, or institutional culture. The result is variable documentation quality, inconsistent structure between clinicians, and a significant proportion of after-hours time spent on notes rather than rest.

AI tools, used correctly, can reduce the cognitive overhead of that final stage. But "used correctly" is doing a lot of work in that sentence.

Why generic AI prompts don't work for psychiatry

The internet has no shortage of prompts like "write me a clinical progress note for a patient with depression." Run that through any current AI assistant and you'll get something that looks plausible from a distance and falls apart on clinical inspection.

The problem is structural. Psychiatric documentation follows conventions that aren't part of general medical or general writing training data. Specifically:

• Mental state examinations have a defined sequence. Appearance, behaviour, speech, mood, affect, thought form, thought content, perception, cognition, insight, and judgment. A generic prompt will either omit domains, conflate them, or render them in an order that signals to any consultant that this wasn't written by a trained clinician.
• Risk documentation in psychiatry is legally significant. A generic "describe the risk" prompt produces narrative prose. What's needed is a structured record of protective and risk factors, the clinical reasoning used to weigh them, and the plan with rationale. The structure matters because it's what gets examined when something goes wrong.
• Psychiatric formulations are not summaries. A generic AI, asked to formulate a case, will typically produce a biopsychosocial list. A proper formulation integrates predisposing, precipitating, perpetuating, and protective factors, and ties them to a hypothesis about why this person is unwell in this way at this time. That requires a prompt that imposes the right framework, not just asks for a "formulation."

In short: the output quality of an AI documentation prompt is almost entirely determined by the structure of the prompt itself. Vague in, vague out.

What makes a good psychiatric prompt

Three principles distinguish prompts that work from prompts that produce generic noise:

1. Structure comes before content

A good prompt doesn't ask the AI to "write a progress note." It specifies the headers, the order, the expected length of each section, and in some cases the exact phrasing conventions. This is how you ensure the output maps onto your EMR fields, your department's template, and the clinical standard expected of your grade.

The prompt is essentially a scaffold. The AI fills in the clinical content you provide; it doesn't invent clinical reasoning. That distinction matters.

2. De-identification is built in, not bolted on

The correct workflow is: you provide de-identified clinical information, the prompt transforms it into a structured document, and you review and edit the output before it goes anywhere. No identifiable patient data should ever be entered into a consumer AI tool.

A well-designed prompt explicitly instructs the AI to work from anonymised clinical notes you provide, and reminds you at the point of use. This is a guardrail, not just a disclaimer.

3. Clinical judgment stays with the clinician

The AI produces a draft. You read it, amend it, and sign it. The AI does not assess risk. It does not make diagnoses. It does not determine capacity. It reformats and structures the information you give it. Prompts that are designed around this principle produce output that is easier to edit and sign off because the AI hasn't tried to insert clinical reasoning it doesn't have.

Practical examples: prompt fragments that work

The following are opening structures from well-designed psychiatric prompts. These are not complete prompts, but they illustrate the structural principle. In each case, the clinician provides de-identified clinical material; the prompt shapes the output.

Progress note opening structure

The bracketed instruction at the end is doing significant work: it prevents the AI from filling in domains with plausible-sounding but fabricated content, which is one of the most common failure modes in AI clinical documentation.

MSE documentation approach

Note that the prompt explicitly asks for negative findings to be stated rather than omitted. In psychiatric notes, documenting that suicidal ideation was assessed and not present is clinically and legally significant. A prompt that doesn't enforce this will silently omit it.

Clinical safety considerations

Using AI for clinical documentation requires a clear-eyed view of what it can and cannot do. The following are not theoretical concerns.

De-identification is non-negotiable. Consumer AI tools (ChatGPT, Claude, Gemini) are not accredited health information processors. Do not enter patient names, dates of birth, MRN numbers, or any other identifying information. The correct workflow is to write de-identified clinical notes for yourself first, then use those as input to the AI prompt. This adds a step but it is the step that keeps you on the right side of your privacy obligations and your hospital's acceptable use policy.

The AI cannot assess. It has no access to the patient, no clinical training, and no accountability. It will produce fluent text. Fluent text is not the same as accurate text. Every AI-generated note must be read, corrected where needed, and signed off by the clinician who conducted the assessment. The note is yours. The AI is a drafting tool.

Hallucination is a real risk. Current AI models can confidently produce plausible-sounding clinical content that was not in your input. The prompt design described above mitigates this by instructing the model to write [not assessed] rather than invent, but it does not eliminate the risk. Read the output.

Check your organisation's policy. Some hospitals have explicit policies on AI use in clinical documentation. Some electronic medical record systems have their own AI integrations that may be the appropriate route. Know your institution's position before adopting any external AI tool for clinical work.

What this looks like in practice

A realistic workflow for a registrar doing an inpatient review: see the patient, write brief de-identified clinical notes on paper or in a draft document, then run those through a well-designed prompt to produce a structured draft. Review the draft, amend anything that's inaccurate or missing, paste into the EMR, and sign. In practice, this can save meaningful time on the formatting and structural scaffolding while preserving clinical accuracy because you are still reading and approving everything.

The efficiency gain isn't primarily about word count. It's about the cognitive shift from "I need to construct a structured document from scratch after a difficult consult" to "I need to review and correct a structured draft." For many registrars, particularly late in a shift, that difference is significant.

Where this fails is when clinicians skip the review step, or when they enter more detail than is needed and trust the output uncritically. The prompt design helps, but the clinical habit matters too.