De-identify client notes
Strip client identifiers from any text while keeping the clinical meaning. Use this before you paste anything into an AI assistant like ChatGPT or Claude, or whenever you need a de-identified version for supervision or teaching.
Do this before anything else. Never paste identifiable client data into a general AI assistant. Under the DPDP Act, client identifiers are sensitive. The safest practice is to de-identify at the source.
How to use it
- 1Copy the prompt above, or download it as a file.
- 2Open your AI assistant (Claude, ChatGPT, Gemini, or any LLM).
- 3Paste the prompt, then paste the text you want de-identified after it.
- 4Read the output and the residual-risk notes before you use or share it.
The prompt
Copy it, or download it as a file to keep or to save as a reusable prompt in your assistant.
You help me remove identifying information from clinical text while keeping the clinical meaning intact. You are a drafting aid, not a clinician. When I paste text: 1. Replace every direct identifier with a neutral role label: names of client, family, partners, colleagues, clinicians become [client], [partner], [mother], [child, age 8], [colleague]. Remove phone, email, address, ID/registration/insurance numbers. Turn employer/school into [employer]/[school]. Generalise exact dates of birth and full addresses to what is clinically needed (e.g. "late 20s", "a city in Kerala"). 2. Watch for indirect identifiers: a combination of small details that together identify someone (rare job + rare condition + small town), unusual public roles or events, named institutions. Generalise or flag these. 3. Preserve everything clinically meaningful: age band, presenting problem, symptoms, timeline, formulation, interventions, risk, plan. Never soften or drop a risk signal while de-identifying. Names go, risk stays. 4. Never re-identify or guess a name you removed. Return: the de-identified text, then a short list of what you changed and any residual re-identification risk you still see. Remind me to review before sharing, and that the safest practice is to de-identify at the source rather than paste raw identifiers into any general AI assistant. Here is my text:
What it looks like
Rahul Menon, 34, software engineer at Infosys Kochi, came in on 12 June after a panic attack at work. His wife Anjali drove him.
After[client], mid-30s, works in IT, presented after a panic attack at work. Accompanied by [partner].
Client is a well-known local news anchor in Thrissur, single mother of a 7-year-old, presenting with insomnia and low mood for 3 weeks.
After[client], single parent of a [child, age 7], presenting with insomnia and low mood for 3 weeks. (Residual risk flagged: the original public role was highly identifying.)
This is a scaffold, not a guarantee. Review the output before you use or share it, and check that no combination of remaining details narrows down who the person is.
Frequently asked questions
Only with de-identified text. Never paste a client's name or identifying details into a general AI assistant. Under India's DPDP Act, client identifiers are sensitive personal data, so de-identify first, then use any assistant.
It removes names, contact details, IDs and obvious identifiers, and flags indirect ones. It is a scaffold, not a guarantee, so always review the output before you share it.
Yes, that is one of the main uses. Once identifiers are removed and you have checked the result, the text is much safer to use in supervision, teaching or peer discussion.
A drafting aid for qualified professionals, not a diagnosis, clinical decision, or legal advice. Never paste identifiable client data into a general AI assistant. If you or someone you are with is in crisis, contact a crisis helpline right away.