Optometry Referral Letter Template: Draft With AI, Then Check

Write the referral from structured fields (reason, findings, history, medicines, what the patient was told, what you are asking for), give those fields to an AI tool as placeholders or approved-tool data, ask for a concise draft, then check laterality, every number, urgency wording and patient identifiers yourself before the letter goes anywhere.

Most referral templates online are blank forms, and a blank form is half the job: the hard parts are getting your findings into clear sentences quickly and being certain nothing wrong survived the drafting. The method below gives you both a template and a repeatable check, and it keeps the clinical judgement, above all urgency, with you.

Follow me on Instagram@sagnikteaches

The template fields, in the order a reader needs them

A GP, ophthalmologist or hospital eye clinic reads dozens of letters a day. They want the ask first and the evidence second. Use these headings every time so any gap shows.

Connect on LinkedInSagnik Bhattacharya

1. Header. Practice name and contact details, date, recipient name or department, patient full name, date of birth and any identifier used locally.
2. Urgency line. Written by the clinician, in the words your referral pathway uses. Never drafted by AI.
3. Reason for referral. One or two sentences: what you found, which eye, and what you are asking the recipient to do.
4. Presenting history. The symptoms, onset and duration in the patient's account.
5. Examination findings. Visual acuities, intraocular pressures, anterior and posterior segment findings, fields or imaging, each with the eye specified and the date.
6. Relevant history. Medical and ocular history, medicines, allergies, family history where relevant.
7. What the patient has been told. And whether they agree to the referral and to sharing information.
8. Request. Assessment, opinion, treatment or shared care.
9. Enclosures. Images, scans, visual field printouts and prior results.
10. Sign-off. Name, professional registration number and contact route for questions.

Two versions of the same structure follow. The shape is the same; the vocabulary and routes differ.

Subscribe on YouTube@sagnikteaches

Version A: referral to a GP

This is for a finding that needs a systemic check or a medicines review, for example signs that suggest blood-pressure or glucose problems, or a medicine with eye effects. Invented example, with every clinical detail made up for illustration:

Dr [GP name], [practice name]
Re: [Patient name], DOB [date]   Date: [date]

Urgency: [clinician's wording]

Reason for referral
I would be grateful if you could review [patient] for possible [systemic concern] after
findings at a routine eye examination on [date].

Findings
Right eye: [finding]. Left eye: [finding]. Visual acuity [R] / [L] with correction.
Intraocular pressures [R] / [L] mmHg at [time].
Imaging: [attached / not taken].

History
[Brief relevant history]. Current medicines as reported by the patient: [list].

Discussed with patient
I explained [what you said]. The patient agrees to this referral and to my sharing
this information with you.

Request
[Review / blood test / medicines review]. Please let me know the outcome so that
I can plan eye follow-up.

[Name], [qualification], [registration number], [contact]

Version B: referral to an ophthalmologist

Use this for a suspected eye condition that needs a specialist's opinion. Again every value is invented:

[Ophthalmologist name], [clinic or hospital eye service]
Re: [Patient name], DOB [date]   Date: [date]

Urgency: [clinician's wording]

Reason for referral
Suspected [condition] in the [right / left] eye. I would value your assessment and
advice on management.

Findings (examination date [date])
[Right / Left] eye: [specific findings, measurements].
Fellow eye: [findings].
Visual acuity: R [ ] / L [ ]. Intraocular pressure: R [ ] / L [ ] mmHg.
Imaging / fields: [attached; summary].

History
Symptoms: [onset, duration]. Ocular history: [ ]. Medical history and medicines: [ ].
Family history: [if relevant].

Discussed with patient
[What was explained; agreement to referral and sharing information].

Enclosures
[Images, scan, field printout].

[Name], [qualification], [registration number], [contact route]

Notice that both versions have the eye named in two places: in the reason and in the findings. Redundancy on laterality is deliberate, and it gives your later check something to compare.

Choosing a tool before any patient data touches it

Drafting from a template with placeholders needs no patient information, and almost any assistant will do. Drafting from real findings is different, and the tool choice matters.

• Placeholders only. Use any assistant to improve the template and wording. Treat everything you type as visible to the provider.
• Real findings. Use a product that your regulator, privacy law and your own policy allow for patient information, with the right agreement signed. In the United States, that means a business associate agreement; elsewhere it means a data-processing arrangement that fits local law. OpenAI's help page for ChatGPT for Clinicians, for example, lists referral letters among the drafting uses for verified US clinicians, and lets an individual clinician sign a BAA inside the product. Which OpenAI product fits which practice sets out what is and is not covered.
• Your own record system. Some practice systems include built-in letter drafting. Ask the vendor how it handles patient data, and test with invented patients first.

If you use an AI note-taker for exams, the structured note is the natural input to a letter. The scribe test plan for optometrists explains how to check that notes are accurate before anything is built on them. And if you tell patients you use AI, the consent wording tutorial gives plain language for it.

The drafting prompt

Paste this with the fields filled from your own notes. In practice, put real values in only an approved tool; build and test with invented ones.

Draft a referral letter from an optometrist to a [GP / ophthalmologist].
Use ONLY the facts below. Do not add, infer or estimate any clinical finding,
measurement, date or diagnosis. Where something is missing, write
[MISSING: what is needed]. Use the heading structure: Reason, Findings,
History, Discussed with patient, Request, Enclosures. Under 300 words.
Leave the urgency line exactly as "Urgency: [CLINICIAN TO COMPLETE]".
Write laterality in words (right eye, left eye), never as R/L alone.
Keep a neutral professional tone and do not state a diagnosis as certain
if I wrote "suspected".

Facts:
Patient: [initials only here; full details added by me]
Reason: [text]
Findings: [text, each with the eye and date]
History: [text]
Told patient: [text]
Request: [text]
Enclosures: [text]

Two details matter. "Do not add, infer or estimate" gives the model less room to fill gaps with plausible but invented numbers. "Write laterality in words" removes the habit of abbreviating, which is where left-right swaps hide. The [MISSING] markers turn omissions into visible tasks, as happens in AI-drafted prior-authorisation letters.

A worked example, with the errors found

An illustrative case. A 58-year-old patient attends for a routine exam. The optometrist finds a raised pressure in the right eye and an optic disc appearance she wants a specialist to assess. Her structured note, written with invented values, says: intraocular pressure right 29, left 18 mmHg; right disc cupping asymmetric; no pain; family history in a parent; no medicines affecting the eye; patient told; agrees to referral.

The model returns a clean 240-word draft. Her checks find four problems:

• The draft says "pressure of 29 mmHg in the left eye". Laterality swapped. This is why the prompt says to write eyes in words, and why you compare against the note.
• It writes "urgent referral" in the reason line, although the urgency line was meant to be hers. She deletes it and writes her own wording.
• It adds "the patient is at high risk of glaucoma", which she did not say. She changes it to "suspected raised pressure with asymmetric cupping", matching her note.
• It lists "no known allergies" when she did not record allergies. She replaces it with [MISSING], checks the record, and fills in the answer.

None of these errors is exotic. They are typical of fluent, confident drafts, which is why the check below is not optional.

The pre-send checklist: ten checks, two minutes

Run these in order on every letter. Print them and keep them by the screen.

1. Patient identity. Name, date of birth and identifier match the record. A wrong patient is the worst error, so read these first.
2. Recipient. Right clinician or department, right address or secure channel.
3. Urgency line. Written by you, matches your own judgement, and no other urgency language appears in the body.
4. Laterality. Every mention of a finding names the correct eye. Compare each with the exam record.
5. Numbers. Every acuity, pressure and measurement matches the record, with units and date.
6. Nothing invented. No diagnosis, risk statement, history item or medicine that is not in your notes. Search the draft for "likely", "high risk", "no known" and "no history".
7. Consent. The patient agrees to the referral and to sharing information, and the letter says so.
8. Enclosures. What the letter says is attached is attached.
9. Tone. "Suspected" stays "suspected". "I would be grateful" has not become "please arrange".
10. Sign-off. Your name, registration number and a working contact route.

Read the letter once as the recipient. If you were the GP or the clinic triage nurse, would you know in the first two lines what is being asked and how soon? If not, move the ask higher.

Same structure, different vocabulary

The template travels between healthcare systems with small changes. Keep one master and a short table of local terms, so staff do not rewrite the whole letter each time the destination changes.

Check each route's own requirements before you finalise the master; some pathways accept only a particular form, and a perfect letter in the wrong format can sit in a queue. This is a good point to ask a colleague at the receiving clinic what makes a referral easy for them to triage. A ten-minute conversation tells you more than a dozen online templates.

Three more cases, described briefly

A gradual-onset cataract. A patient in her seventies finds night driving difficult. The optometrist records reduced acuity in both eyes with a lens change visible on examination, and the patient wants to discuss surgery. The letter is routine in tone, with a clear request ("assessment for cataract surgery"), a short functional impact line in the patient's own words, and enclosures of the best-corrected acuities. The AI draft tends to add reassurance ("this is common and treatable") that does not belong in a clinical letter, so the check for invented statements strikes it.

Sudden flashes and floaters. Here the urgency line carries the weight, and the pathway may call for a phone call before any letter is sent. The letter is a record of what was found and said; the clinician sets the timing and route. If you use AI at all for this case, use it only to tidy a letter after the call, never to decide what to do.

A child with a persistent squint. The letter needs the parent or guardian's agreement, age, and the findings with correct eyes. Because it involves a minor, tighten your tool choice and your sharing checks. The model may want to say "amblyopia" because it is common in similar letters; unless that is your finding, delete it.

A rambling letter, rewritten

AI is also useful for tidying a letter you wrote in a hurry. The prompt is simple: "Rewrite this referral letter into the template headings. Keep every fact. Do not add anything. List any fact that appears twice or that seems to conflict." An illustrative before and after, using invented content:

Before: a single 190-word paragraph that mixes symptoms, a pressure reading, a family comment and the request, with "R" and "L" abbreviations and the ask in the last line. After: a 150-word letter whose first line states the ask, followed by labelled findings for each eye, with the family history moved under History and the patient's agreement stated once. The "conflict" list caught that the pressure appeared as 24 in one place and 26 in another, which the clinician resolved from the record. That catch alone makes the exercise worthwhile.

Testing the prompt on five invented cases

Before you trust the prompt for real letters, run it on five invented cases and score each output against the checklist. Choose cases that stress it:

1. One with a missing field, to see whether it writes [MISSING] or invents.
2. One where the abnormal finding is in the left eye and the history mentions the right eye.
3. One with two similar measurements on different dates.
4. One with a "suspected" diagnosis, to see whether it stays suspected.
5. One with a very short note, to see whether it pads.

Score each out of ten against the pre-send checklist. If any case fails on laterality or invention, adjust the prompt and rerun all five. When all five pass twice in a row, you have a prompt you can use with confidence, and an evidence file to show a colleague or regulator if asked how you check AI output. Retest after a model or tool update, because behaviour can change even if the prompt does not.

Wording to keep out of the letter

• Reassurance and prediction. "This is nothing to worry about" and "the outlook is good" are opinions the recipient did not ask for.
• Blame or speculation about earlier care. Stay with facts.
• Abbreviations the recipient may not share. Spell out terms that differ between professions.
• Padding. Courtesy lines add length without information. One sentence of thanks is enough.
• Unrequested extras. Extra history the recipient does not need conflicts with the sharing-only-what-is-needed principle.

Keeping the template library tidy

Store the master template, the prompt and the checklist in one shared place with a version date and an owner. When a recipient asks for a change, update the master and note it. When you find a better prompt line, test it on the five cases before replacing the old one. Archive old versions so you can show which wording was in use on a given date. A template library that nobody owns slowly splits into five personal copies.

What the time saving looks like, and where it goes

An illustrative sum for a practice that sends 12 referral letters a week. Writing each from scratch takes about 12 minutes, so 144 minutes a week. With structured notes, the draft and the ten-point check, a letter takes about 6 minutes: 2 to enter or confirm the fields, 1 for the draft, and 3 for the checks and the urgency line. That is 72 minutes a week instead of 144, a saving of 72 minutes, or about 55 hours over a 46-week working year. Do not spend all of it elsewhere. Put 10 minutes a week into the error log and the monthly prompt review, because that is what keeps the saving from turning into a quiet drift in quality.

If the saving is smaller than this in your first month, check where the time goes. The usual cause is incomplete notes, which force the model (and you) to chase missing details. Improving the structure of the exam note usually helps more than improving the prompt.

A second pair of eyes for the first month

For your first twenty AI-assisted letters, ask a colleague to read a sample of five against the original notes, without telling them which draft was AI-assisted. Ask one question: "Is anything in this letter that is not in the notes?" Their answer gives you an honest error rate and builds shared confidence in the method. If they find nothing across five, good; if they find an invented line, show them the matching checklist item and tighten the prompt. After the first month, move to a spot check of one letter in ten.

Keeping records and learning from slips

Save the final letter and the structured note it came from to the patient record, and not the chat. Keep a record of what was sent, when and how. If a recipient replies with a correction or a question that shows something was unclear, note it. After twenty letters, review the questions you were asked: they tell you which field of the template needs a clearer prompt.

A small error log helps. A one-line entry for each catch ("left/right swapped in draft; caught at check four") gives you evidence about which checks are worth the two minutes and which prompt line did the work. If the same error appears more than twice, change the prompt. If a certain type of letter keeps needing heavy edits, write it by hand and do not use AI for it.

Patient-facing versions

Patients often want a plain-language explanation of why they are being referred. The same structured note can produce a short, calm explanation: what was found, what the specialist will check, what to expect, who to contact if symptoms change. The clinician approves it, it avoids any prediction of outcome, and it uses the same facts as the letter so nothing contradicts it. For ideas on plain wording about eye test results, the tutorial on explaining lens options and results with AI has examples.

What stays with the clinician

• Deciding that a referral is needed, and to whom.
• Setting urgency and the safety-netting advice to the patient.
• Interpreting findings, and anything labelled a diagnosis.
• Obtaining and recording consent to share.
• Signing the letter.

AI does the typing: sorting facts into a structure and phrasing them clearly. If you want to talk through where drafting tools fit in your practice and where they do not, that is what my AI implementation consultation covers, via the contact page. For a related patient-facing job, see recall letters and review replies with ChatGPT.

Questions optometrists ask about AI referral letters

Should the AI decide how urgent the referral is?

No. Urgency is a clinical judgement and the wording that signals it should come from the clinician. Leave the urgency line blank in the prompt, then type it yourself after drafting, and read the final letter to check the model has not softened or inflated it elsewhere.

Can I paste exam findings into ChatGPT or Claude?

Only if the tool and plan are suitable for patient information and your regulator and local privacy rules allow it. A consumer chat account is not a safe default. Use placeholders and invented data to build the template, and enter real findings only in a product with the right agreement.

Is a template enough, or should every letter be written fresh?

A template with fixed headings is the right base because it makes omissions obvious. The reasoning in each letter should still be the clinician's own. Use AI to turn your structured notes into clear sentences, not to supply findings or an opinion.

How long should a referral letter be?

Short enough to be read on one screen or one page: reason, key findings, relevant history, what you have told the patient and what you are asking for. Anything the recipient needs but cannot fit should be attached as a measurement, scan or report instead of described at length.

Further reads

• AI in an Optometry Practice: A Week-by-Week First-Month Plan — A week-by-week plan for bringing AI into the practice.
• AI Receptionist for Optometrists: Eye Test Bookings and Recalls — Bookings and recalls with an AI receptionist.
• AI Scribe Security: Audio, Retention and BAA Vendor Questions — Vendor questions on audio, retention and BAAs.
• AI Usage Policy for a Medical Practice: Template and HIPAA Rules — A staff AI-use policy template for a medical practice.
• Optometry Insurance Verification: A Front-Desk Workflow With AI — A front-desk routine that sits before the visit.
• Eight Practical AI Uses for an Independent Optician — Eight practical AI uses for an independent optician.
• Dental Practice AI Mistakes: Consent, Data, and Over-Automation — Consent, data and over-automation errors in a clinical setting.
• AI for Optometrists: Contact Lens Aftercare and Reorder Messages — Draft contact lens aftercare and reorder message sequences with AI, with stop rules and a route for discomfort replies to staff.
• AI Patient Education Materials: A Fact-Check and Reading-Level Checklist — A clinic checklist for AI-drafted patient handouts: sources first, clinician sign-off, a reading-level test and a version log.
• GDPR in an Optical Practice: Which AI Tools May Touch Records — A scored 12-point checklist for vetting any AI tool before it touches eye-test records, with four tools worked through and the paperwork behind each.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: clinical-administration practice for eye-care letters; OpenAI's ChatGPT for Clinicians help page (checked 4 November 2026) for the tool-choice note. All clinical values in examples are invented. Check your own regulator's record-keeping and referral guidance.