Lesson objective
Write prompt constraints that keep a draft administrative and out of clinical territory, and keep licensed code descriptors out of both the prompt and the output.
Plain-language explanation
Why this matters now
A dental office prompt can be perfectly built and still produce the wrong kind of text. Asked to write a friendly recall message, a model may add that "regular cleanings prevent gum disease." Asked to summarise a benefits question, it may describe what a procedure involves. Neither sentence was requested, and both move the draft into clinical territory that belongs to the licensed dentist.
A second boundary is less obvious: the text of the dental procedure code set. It is licensed material, and a model's memory of it is not a source.
This lesson turns both boundaries into prompt constraints, and into a check the reviewer runs on every output.
A question to sit with
A coordinator asks an assistant to explain a benefits denial to a patient in plain words, and pastes the code descriptor from the practice's manual "so it understands." The draft is clear, kind and mentions what the procedure treats. What has gone wrong, twice?
What goes wrong
Unrequested clinical content. The model adds a benefit of treatment, a consequence of waiting, a description of a condition or a reassurance about pain. Each is clinical information the dentist has not approved.
Urgency by tone. Words like "soon", "important" or "before it gets worse" imply an urgency judgement that no administrative staff member can make.
Implied treatment. "Once your crown is approved" implies the dentist has decided on a crown and that it will happen.
Pasting licensed text. Descriptors from the code manual are copied into a tool without anyone checking whether the practice's licence allows it.
Treating recall as a source. The model is asked what a code "means" and its answer is used as if it came from the manual.
The practical answer
Part one: constraints that hold a draft on the administrative side. Put these in every patient-facing or benefits prompt, in plain words. They are this programme's recommendation; no dental prompt research has tested them, which is why the reviewer's check follows.
| Constraint | Wording you can adapt |
|---|---|
| No clinical content | Do not describe any condition, finding, symptom, risk, benefit of treatment or consequence of delay |
| No urgency | Do not say or suggest how soon anything should happen, except dates given in the source |
| No implied treatment | Do not suggest that any treatment, prescription or referral has been decided or will follow |
| No clinical reassurance | Do not comment on pain, comfort, healing or outcomes |
| Route clinical questions | If the source contains a clinical question, reply only that it will be passed to the dentist |
| Supplied material only | Use only the labelled sources; state what is missing |
The reviewer's boundary check. A constraint reduces the problem; it does not remove it. The reviewer reads every output for the five warning signs and deletes or routes what they find.
| Warning sign | Example in a draft | Action |
|---|---|---|
| A condition or finding named | "to treat the decay" | Delete; the patient's clinical information comes from the dentist |
| A benefit or risk of treatment | "cleanings help prevent gum disease" | Delete, or route to the dentist if the practice wants such text approved |
| Urgency | "please book soon to avoid problems" | Replace with the scheduling facts only |
| Implied decision | "when your implant is done" | Rewrite neutrally, or route to the dentist |
| Clinical reassurance | "this is a routine, painless procedure" | Delete |
A draft that needs clinical content to make sense goes to the licensed dentist before it goes anywhere else. The dentist owns any communication that conveys clinical information.
Part two: licensed code text stays out.
What that means for a prompt:
- Do not paste descriptors from the manual into a tool unless the practice has confirmed its licence allows it. The office manager or practice owner makes that call.
- Do not ask a model to reproduce or explain descriptors. The ADA page does not publish them, and a model's recollection of one is not a source.
- Do not ask a model which code fits a service. Choosing a code is not part of this programme; the billing lead works from what the dentist documented, and a coding question goes to the billing lead and, where needed, the dentist.
- Check the output. If a draft contains anything that reads like a code descriptor, delete it and check where it came from.
- Use the current manual as the source. The ADA provides free coding assistance to members and to non-members who bought the current manual.
What a benefits prompt can say instead. Refer to the service the way the practice's own ledger or the dentist's signed record shows it to staff, and tell the model to describe the benefit position (covered percentage, remaining maximum, frequency limit as quoted from the plan) without describing the procedure.
The Giannakopoulos study (2023) is a reminder of why clinical explanation stays out: even on questions written and graded by dental faculty, every model sometimes gave inaccurate or outdated content without sources.
Remember this
- Every patient-facing or benefits prompt carries the no-clinical-content, no-urgency and no-implied-treatment constraints.
- The reviewer still reads for the five warning signs, because constraints reduce the problem rather than remove it.
- Clinical explanation comes from the licensed dentist, never from a draft.
- CDT descriptors are licensed; do not paste them without checking the licence, and never treat a model's version as a source.
- No prompt asks which code fits a service; that question goes to the billing lead working from the dentist's record.
Dental practice example
Cedarbrook Family Dentistry, a fictional practice, has a saved prompt for explaining benefits denials to patients. A new coordinator, Alana Brooks, uses it for a fictional patient, Mr. Desmond Rourke, and adds the code descriptor from the practice's manual "for context." The draft reads well but says the service "treats the infection so it does not spread" and asks him to "reschedule soon so the tooth is not lost." The office manager, Tessa Morrow, catches it at review. She removes both sentences: the first describes a condition and the second is an urgency judgement, neither of which the dentist approved. She also removes the pasted descriptor from the prompt history and asks the practice owner, Dr. Imran Qadir, whether the practice's licence allows manual text in the assistant; until he answers, the rule is no descriptors in any prompt. The rewritten prompt supplies only the payer's denial reason as quoted in the explanation of benefits, the plan year, the remaining annual maximum of $640 and the service as the ledger shows it to staff, with the no-clinical-content and no-urgency constraints added. The new draft explains the denial reason, the maximum and how to reach the billing office, and says that questions about his treatment will go to Dr. Qadir. Tessa checks the $640 against the ledger and the denial wording against the explanation of benefits, then approves it. Because the patient's call earlier had included a question about whether he needed the treatment at all, she also puts a note in Dr. Qadir's queue. Both removals are recorded on the use record.
Responsible-use warning
Do not treat readable as harmless. The clinical sentences a model adds to an administrative draft are usually kind and plausible, which is why they slip through. A benefit of treatment, a consequence of waiting or a reassurance about pain is clinical communication the dentist has not approved. And do not paste code manual text into a tool, or ask a model to supply or explain a descriptor: the text is licensed, the model's version is not a source, and choosing a code is never the prompt's job.
Human-review requirement
The office manager or named administrative owner reads every patient-facing or benefits draft for the five warning signs and removes or routes anything clinical. The licensed dentist owns and approves any message that conveys clinical information, and receives any patient question about need, urgency or treatment. The billing lead owns anything that touches a code, working from the dentist's documentation and the current manual, never from a model. The practice owner decides whether the practice's licence allows any manual text in any tool, and that decision is recorded.
Practical exercise
Use a fictional practice, fictional patients and fictional plan documents only. Do not copy any real code descriptor.
- Write a fictional recall message prompt with no boundary constraints.
- Write a fictional draft it might produce that contains at least three of the five warning signs.
- Mark each warning sign and write the action the reviewer would take.
- Add the six constraints from this lesson to the prompt in your own words.
- Write a fictional benefits denial prompt that refers to the service the way a ledger would show it to staff, without any descriptor.
- List three things the prompt may say about the benefit position, each tied to a quoted plan sentence.
- Write the sentence the draft uses when a patient's message contains a clinical question.
- Write a short note to your practice owner asking whether the practice's code manual licence allows any manual text in the assistant.
- Write one line explaining why a model's explanation of a code is not a source.
- Name who reviews the draft, who approves clinical content and who answers coding questions.
- Record the boundary check and any removals on the AI Use Record in your workbook, and note which constraint, if any, failed to prevent each one.
Knowledge check
-
Name the five warning signs a reviewer reads for.
Answer guidance: A condition or finding named; a benefit or risk of treatment; urgency; an implied treatment decision; and clinical reassurance about pain, comfort, healing or outcomes.
-
Why is a model's explanation of a CDT descriptor not usable?
Answer guidance: The CDT manual is copyrighted and licensed, the ADA page does not publish descriptors, and a model's recollection is not a source. The current manual is the source.
-
What makes the CDT Code binding in practice, and who maintains it?
Answer guidance: HIPAA adopts it for dental services at 45 CFR 162.1002. The ADA maintains it, with its Code Maintenance Committee deciding changes; the ADA is a maintainer, not a regulator.
-
What should a draft say when a patient's message contains a clinical question?
Answer guidance: Only that the question will be passed to the dentist. The draft gives no clinical answer, and the question is routed to the licensed dentist.
Sources and further reading
- American Dental Association — The Code on Dental Procedures and Nomenclature (CDT Code). Accessed 2026-09-15.
- National Institute of Standards and Technology — Artificial Intelligence Risk Management Framework: Generative Artificial Intelligence Profile, NIST AI 600-1 (2024-07-26). Accessed 2026-09-15.
- Journal of Medical Internet Research — Evaluation of the Performance of Generative AI Large Language Models ChatGPT, Google Bard, and Microsoft Bing Chat in Supporting Evidence-Based Dentistry: Comparative Mixed Methods Study (2023-12-28). Accessed 2026-09-15.
- AI Coalition Network — Certification Educational Standard: Human Review of AI-Assisted Professional Work. Accessed 2026-09-15.