What I may and may not do

Built to be printed and handed to someone — a colleague, a compliance officer, an IRB. It carries its own caveats and its own key, because paper cannot link anywhere.

This is a teaching aid, not compliance or legal advice. Decisions about patient data belong to your organization, its counsel, and its Anthropic contact. Anthropic's policies and products change frequently — confirm current specifics with your Anthropic contact before relying on any line of this sheet for a compliance decision.

Every ✓ and ⚠ rule below carries the date it was last checked against an Anthropic page; the newest on this sheet is 2026-09-19, and none is later. The 5 rules with no date were never checked against an Anthropic page. 16 Anthropic primary pages cited in all.
Counted as: distinct Anthropic-published pages cited by the 18 rules on this sheet. 5 lines below are marked ✎ our practice — practice rules from this curriculum and the working group's own document, NOT Anthropic product facts, and not part of that count.

How to read the marks:

Confirmed checked against Anthropic's own page on the date shown
Documented not checked against an Anthropic page — the group's document, or our own practice
Stale risk treat as a question, not a fact
compliance a patient-data decision could hang on this
our practice a practice rule of ours, not an Anthropic product fact

You may

Give Claude one dedicated working folder. Never broad access to your desktop or documents.

Confirmed · 2026-07-25
Keep credentials in a secrets manager or environment variables — outside any folder Claude can read.

Documented our practice
On a commercial plan, what you type is not used to train models — by default and by contract.

Confirmed · 2026-09-16 compliance
Get a BAA in place before any patient data goes near Claude. Two configurations qualify — a HIPAA-ready API organization, or an Enterprise plan with HIPAA switched on. Most organizations can set either up without a sales call — one that requires a negotiated BAA works with its account team — and coverage is not automatic on the right plan: on the Enterprise path only the Primary Owner can accept the BAA, and switching HIPAA on cannot be undone from organization settings.

Confirmed · 2026-08-03 compliance
Run the de-identification pre-flight before anything is pasted.

Documented our practice
Write a simple Skill without any code — it is a Markdown file saying what the Skill does and when Claude should use it. Do not plan around an interactive builder: the skill-creator helper circulates in write-ups, and no Anthropic page we can find describes it.

Confirmed · 2026-07-27
Run the three checks every time: is the model available, is the feature eligible today, and did you check the output against a source you opened yourself.

Documented our practice

You must not

Never paste a credential into a chat, a shared folder, or an artifact.

Documented our practice
Do not rate sensitive conversations or attach them to bug reports. Feedback can send the related conversation; feedback and bug submissions have a five-year retention policy and can permit training. Other explicit opt-ins can also permit training.

Confirmed · 2026-09-16 compliance
Do not treat an incognito chat as private on Team or Enterprise. It is still retained, still exported to your account Owners, and — on Enterprise — still visible to the Compliance API.

Confirmed · 2026-07-28 compliance
Do not put anything sensitive inside a Skill. Skills sit outside zero-data-retention, and outside HIPAA coverage on the API path.

Confirmed · 2026-07-28 compliance
Do not assume your way of using Claude is covered. Most are never covered for patient data, and on two cloud platforms the decision is not Anthropic's to make.

Confirmed · 2026-07-28 compliance
Never put patient data into a JSON schema. Cached schemas are not protected the way message content is.

Confirmed · 2026-07-25 compliance
Nothing patient- or family-facing goes out without a clinician reading it first — diagnosis, treatment, prognosis, eligibility, what a family should do next. Nothing else goes out without someone who can check it having checked it.

Documented our practice
Do not treat outside files, pages, or email as inert — any of them can carry instructions aimed at Claude rather than at you. Narrow what Claude can read, or narrow what it can do — you do not have to do both.

Confirmed · 2026-07-25 compliance

Ask before you assume

Verify Enterprise account, session type and settings before attesting Compliance API coverage. Local HIPAA-readiness and ZDR sessions are excluded. Audit-log export and OpenTelemetry are separate routes.

Confirmed · 2026-09-16 compliance
If you turn on Cowork monitoring, ask what your collector actually receives. Two Anthropic pages disagree about whether it carries the words your staff typed — which decides whether your monitoring system is now a store of patient information you have to govern.

Stale risk · 2026-09-16 compliance
Ask what your administrators can centrally manage about Skills. Two live Anthropic pages disagree — one says Owners can provision Skills for everyone, the other says Skills cannot be centrally managed.

Stale risk · 2026-09-19 compliance

This sheet summarizes a longer curriculum with the evidence behind every line, at rareanthropic-curriculum.vercel.app. Each rule above is one finding there, with its primary source and, where it has one, the date it was checked. Not affiliated with or endorsed by Anthropic; quotations are from Anthropic's published pages.