The skill behind this guide: Nora - Medical Writer AI Skill. Asks for audience, document type and jurisdiction first, and marks missing data instead of inventing it - $19, one payment, yours permanently.
View the Nora skill →A model can take the first-draft burden off structuring a document, recasting the same trial data for a different reader, and turning clinical language into plain language. It cannot be trusted with a reference, it does not know your approved label, and it has no idea which of your inputs you were allowed to paste. Nothing drafted this way goes to a patient, a regulator or a journal without qualified medical, scientific and regulatory review. This is not medical advice.
Why is citation fabrication the first thing to deal with?
Because in this field a fake reference does not just embarrass you. A model will produce a real-sounding author list, a plausible journal, a volume, a page range and a well-formed DOI for a paper that does not exist - and the output carries no signal that it invented any of it. No hedge, no missing field, no lower confidence.
It does this most often precisely where you most want the support: the specific claim at the hinge of your argument, in the place where the real literature is thinnest. If a well-covered topic is being cited, the model has seen genuine papers and is more likely to name one. If your sentence needs a study nobody ran, you get a study nobody ran.
The adjacent failures are worse in medical writing because they survive a quick look: a real paper with invented page numbers; a real paper whose finding was in a different population, at a different dose, or over a different follow-up than the sentence you attached it to; a resolving DOI pointing at an unrelated article. Every reference gets resolved against the actual source - DOI, PubMed, the journal - before it goes anywhere.
The method that works: never ask for sources
Reverse the order. Have the model mark the claims that need citing and tell you what to search for. Find the sources yourself. Then paste the real source text back in and have it check your draft against what the source actually says - which is the step that catches the more common problem of a real reference stretched past its evidence.
Prompt 1 - mark the claims, supply no references
Prompt 2 - check the paraphrase against the real source
Asks for audience, document type and jurisdiction before it writes a line, then works only from the data you give it - marking anything absent as [DATA TO BE PROVIDED] rather than filling the gap. Every output comes back labelled as a draft for qualified review.
View Nora - Medical Writer AI Skill →What can you actually put into it?
This is a real decision with real rules, and it is decided before you paste, not after. Two separate problems get confused.
Patient data. Under HIPAA, a vendor processing protected health information on your behalf needs a Business Associate Agreement. Consumer chat tiers generally do not come with one, so "I will just paste the case summary" is a decision about your organisation's compliance posture, not a convenience. De-identification also has defined standards - Safe Harbor and Expert Determination - and removing the name is not one of them. Dates, a rare diagnosis, a small site population, an unusual combination of ages and locations: any of these can re-identify a person. Under GDPR, health data is a special category with its own basis, minimisation and retention requirements, and where the processing physically happens matters.
Unpublished trial data. Different rulebook entirely. This is usually confidential to the sponsor under contract, and a confidentiality agreement does not care whether the recipient is a person or a tool. Topline results before disclosure can also be material non-public information for a listed sponsor.
The practical answer: ask what tier and what agreement your organisation has in place, and get the answer before you draft anything real. Learn the workflow on synthetic or already-published data - it teaches you exactly the same thing.
Which documents does this actually work for?
Not evenly. The useful split is not by therapeutic area but by how much of the document is structure and register, and how much of it is regulated claim. The more the document is a rearrangement of data you already hold, the more a model helps. The more it is an assertion about a medicine, the more it becomes a liability you have to police line by line.
| Document | What the model genuinely does | The hard limit |
|---|---|---|
|
Plain-language summary lay results summary, patient leaflet copy |
Register shift, reading level, structure, removing hedging that confuses a lay reader without removing the uncertainty itself | It will quietly upgrade "no significant difference" into reassurance. Every claim compared back to the source text |
|
Manuscript sections introduction, discussion, limitations |
Shape of the argument, the limitations you are underplaying, the reviewer objection you have not answered | References. Never take one from the model. ICMJE also bars it from the author line |
| Clinical study report sections | ICH E3 skeleton, consistent cross-referencing, spotting a section you have not populated | Silent omission of a required item, and any arithmetic re-derived from your tables |
| Promotional and med-ed copy | First pass at tone and flow once the claims are fixed by someone else | It does not know your label. Fair balance under 21 CFR 202.1 is a structural requirement it will not meet by default |
| Clinic letters, case summaries | Dictation cleanup, consistent house structure, the correspondence half of the job | The confidentiality question, decided before you paste. Never coded terminology from memory |
| SmPC, package leaflet, prescribing information | Checking your draft against the required section order | Asked to write one it will blend US and EU formats into a document that is neither |
What does the model not know about regulated content?
It does not know your label. It writes from the published literature and from how drug copy generally sounds, which is why it drifts into efficacy phrasing that goes beyond an approved indication without any sense that it has crossed a line.
In the US, prescription drug promotion sits under 21 CFR 202.1. Material must present a fair balance between risk and benefit information - a structural requirement about how thoroughly each is treated, not a matter of tone - and must not be false or misleading about side effects, contraindications or effectiveness, or omit material facts about consequences of use as recommended. Promotional materials are submitted to FDA on Form 2253 at first dissemination, and the Office of Prescription Drug Promotion issues untitled and warning letters when they are not compliant. A model asked for "engaging" copy will produce benefit-heavy text by default, because that is what most marketing copy looks like.
Structured documents have prescribed contents: ICH E3 for clinical study reports, the SmPC and package leaflet formats in the EU, and reporting guidelines such as CONSORT for randomised trials, PRISMA for systematic reviews and STROBE for observational studies. The characteristic failure is not a section that looks wrong - it is a section that reads perfectly and silently omits a required item.
In the EU, the Clinical Trials Regulation adds a deliverable that is squarely a writing problem: alongside the technical summary of results, sponsors must supply a summary written in terms understandable to a layperson. It is the one document where register is the regulated property, which is exactly where a model is strongest and most dangerous at the same time - it is very good at making text readable, and readable is one step from reassuring.
For journal submission, ICMJE recommendations state that a chatbot cannot be listed as an author, because authorship requires responsibility for the accuracy and integrity of the work. Authors are asked to disclose AI-assisted technologies at submission, describe how they were used, and remain responsible for verifying everything an AI touched. The same discipline is covered from the academic side in our academic writer guide, and reference-checking in the citation manager guide.
What is it genuinely good at?
Recasting. The same dataset has to become a clinical study report section, a manuscript, a clinician brief and a patient leaflet, in four registers, and that translation work is real hours. A model given the numbers and told the audience does the structural and linguistic lift while you keep the judgement.
It is also good at first-pass structure for a document type you write rarely, at flagging where your own draft has drifted in register, and at producing the awkward questions a reviewer will ask before the reviewer asks them.
Prompt 3 - recast for a different reader without adding claims
Prompt 4 - pre-review pass before it reaches your reviewer
Prompt 5 - the reassurance audit on a lay summary
Run this one on any patient-facing text, and run it separately from the rewrite that produced the text. Asking the same conversation to both write and audit gets you a defence of what it just wrote.
For the everyday half of the job - clinic letters, case summaries, discharge and referral correspondence, patient education text - drafted for clinician review before anything is sent. The same confidentiality question applies before you paste.
View Healthcare Documentation Prompt Library →How do you install the Nora skill?
The download is a ZIP with SKILL.md at the root of the archive - not inside a nested folder, which is the usual reason an upload fails. In the Claude desktop app, open Customize → Skills, upload the ZIP, and toggle it on. Claude loads it whenever your request matches what it covers.
Skills are available on the Pro, Max, Team and Enterprise plans, with code execution enabled. On Pro and Max you turn on Code execution and file creation under Settings → Capabilities yourself; on Team and Enterprise that switch sits with an owner in organisation settings. If the Skills menu is greyed out on your work account, that is an admin switch someone has turned off - and in a life sciences organisation it may have been turned off deliberately, which is worth asking about before you route around it. The full walkthrough, including the folder-structure mistake that breaks most first uploads, is in the skill installation guide.
In ChatGPT or Gemini there is no upload step: open SKILL.md, copy the contents, and paste them into custom instructions. You lose automatic triggering and keep the method.
Where does it fail?
- Arithmetic on your data. It will re-derive a percentage, a confidence interval or an incidence rate from your text and get it wrong while sounding certain. Copy every figure by hand from the source table.
- Coded terminology. MedDRA preferred terms, ICD and ATC codes come back plausible and frequently wrong. These are lookups, not language tasks.
- Jurisdictional structure. A US prescribing information document and an EU SmPC are not the same document with different spelling. Asked for "prescribing information" it will blend both.
- Fair balance. Left to its own defaults it writes benefit-forward copy. Balance has to be an explicit instruction and still has to be checked by someone qualified.
- Agreeableness. Ask whether your draft is compliant and you will usually be told yes. Ask what a reviewer would object to instead.
- Audit trail. Chat output has no version history, no reviewer sign-off and no controlled record. In an environment governed by 21 CFR Part 11 or EU Annex 11, the draft has to enter your validated document system, not live in a conversation.
- Confidentiality. This one fails at the moment of pasting. There is no later fix.
Who is this for?
Medical writers, regulatory and clinical teams in pharma and CROs, medical communications agencies, and academic researchers writing studies up. It works in Claude, ChatGPT or any AI chat.
Medical writing sits next to half a dozen adjacent jobs, and the catalogue splits them rather than bundling them into one vague healthcare assistant. Each of these is a separate skill at $19:
- Felicitas - Medical Affairs Writer - the medical affairs and pharmacovigilance side of the same documents
- Adalgisa - Regulatory Affairs Specialist - submission structure and regulatory correspondence
- Ethan - Medical Scribe - encounter documentation rather than regulatory writing
- Sancha - Health Writer - health journalism and consumer-facing explanation
- Thibault - Medical Biller and Coder - the coding half, where lookups beat language
For the wider set - consent forms, clinical summaries, nursing and allied health documents - see the healthcare, science & education skills collection and the health & medical prompt packs. Everything here produces drafts for qualified review, and none of it is medical advice.
In summary:
Never take a reference from a model, settle the confidentiality question before you paste, and treat every output as a draft for qualified medical and regulatory review. For regulatory, clinical and patient-facing documents use Nora - Medical Writer AI Skill ($19); for day-to-day clinic letters and case summaries, the Healthcare Documentation Prompt Library ($9) covers the rest. Both work in Claude, ChatGPT and any AI chat, with a 30-day money-back guarantee.
Nora - Medical Writer AI Skill
One file that fixes audience, document type and jurisdiction before a word is written, works only from the data you supply, and marks every gap as [DATA TO BE PROVIDED] instead of inventing one. Drafts for qualified review. No subscription. Yours permanently.
KissMySkills is a marketplace of 2,000+ AI skills, 300+ prompt packs, 100+ agents & free tools for Claude, ChatGPT & any AI chat.
Related Skill Guides
- How to Use Claude as a Medical Scribe: The Ethan Skill Guide
- How to Use Claude as an Academic Writer: What It Can and Cannot Do
- How to Use Claude as a Therapist's Documentation Aid: The Sophia Skill Guide
- How to Use Claude for Mental Health Documentation: The Benjamin Skill Guide
- How to Use Claude as a Pharmacist's Documentation Aid: The Catherine Skill Guide
- How to Use Claude as an Allied Health Assistant: The Lucas Skill Guide
- How to Use Claude as a Nutrition Coach (and Why Its Calorie Numbers Are Guesses)
