Open-Questions Register

Ask Anthropic to recommend a folder layout for the working group's actual files — conference logistics, registry/biorepository spreadsheets, grant materials — so a session working on one org's task can't wander into another org's sensitive files.

Ask whether Cowork can wall off one project's file access from another's, so a task running in Project A truly cannot reach files granted to Project B.

Ask whether there's an audit trail showing exactly which files Claude read, wrote, or deleted, and how to pull back access once it's no longer needed.

If the working group sets up one shared Cowork "Project" for multiple people, ask how access between those members is actually governed — who can see, add, or remove what.

Before picking a surface for a given task, ask Anthropic to lay out plainly what each surface — chat, Cowork, Claude Code — can and can't reach on your files, so the group picks the right tool instead of guessing.

Ask specifically what happens if a credential accidentally appears in a conversation, upload, or connected file — is it stored anywhere, and for how long, separate from the general retention rules.

Ask whether Claude has any automatic protection against reading or exposing a secrets file that accidentally ends up inside a folder it's been given access to — the practical answer today is "keep it out," but ask if there's a backstop too.

Ask for a concrete key-management routine sized for a small nonprofit team, not an enterprise security department.

Before trusting any retention or training answer, confirm which contract actually governs your account — Commercial Terms (Claude for Work, Education, Gov) or Consumer Terms (Free, Pro, Max: a toggle you must check). Then ask for the exact model-training language in those terms, in writing.

Ask how long "deleted" really takes to mean gone, and whether any exception — like a trust-and-safety review — can extend that window.

Ask for a single, plain comparison of what's kept and for how long across every surface the group actually uses — chat, Cowork, Claude Code, and the API — since the answer isn't one-size-fits-all.

If some members are doing org work from their own personal Claude accounts rather than an organizational plan, ask what that actually exposes and whether it's worth standardizing everyone onto Team or Enterprise.

This is a member-added aspiration as much as a question — the group wants to build its own reusable Skills for recurring, detail-sensitive tasks, with built-in checks for accuracy. Ask Anthropic how to approach this well.

Ask for the precise settings path, for each plan type the group uses, to see and confirm the "Help Improve Claude" toggle's current state — don't assume it's off just because no one turned it on.

Ask Anthropic to help match your organization's actual work — natural history studies, registries, biorepositories, patient-reported outcomes — to one of the two BAA-eligible configurations, with real budget numbers, rather than guessing which is "enough."

We know Skills sit outside Zero Data Retention on the API. We do not know their status under an Enterprise BAA, because the document that would say so is one we cannot read. Ask your Anthropic contact for the Skills row in the HIPAA Implementation Guide — or for access to the Guide itself.

If your organization signed a BAA for the Claude API before December 2, 2025, it covers API usage only — it does not extend to a HIPAA-ready Enterprise plan. Find out which side of that date your agreement falls on, and get written confirmation of what it covers now.

Ask for a concrete, practical walkthrough — steps, how long it takes, and what it costs — for actually getting a BAA in place, not just the fact that it's possible.

If your organization isn't on a BAA-covered configuration, ask for a concrete pre-flight checklist for stripping identifying details before anything patient-related goes into Claude at all.

"We think it's de-identified" isn't the same as Anthropic (or HIPAA) agreeing — ask what specific threshold or standard Anthropic expects before treating data as no longer PHI.

A signed BAA doesn't automatically cover every feature — ask for a feature-by-feature list of what's actually safe to use with patient-related data on your specific plan.

A BAA with your registry vendor does not automatically extend to your own direct use of Claude — ask Anthropic to confirm whether these are always separate agreements.

Ask for the practical bottom line: what concrete steps raise the group's security posture? And under exactly what conditions — which plan, which configuration, which feature — is it ever appropriate to type sensitive information into Claude at all?

Before building from scratch, ask whether other nonprofits in similar situations have already built — and would share — a Skill the group could adapt.

Ask whether Anthropic would run a hands-on session building a real Skill together with the group, rather than the group learning solely from documentation.

Ask exactly how a Skill built by one person gets shared with teammates or other organizations — is it a simple export and import, or does it require an Enterprise/Team admin to deploy it org-wide?

Even an organization with a Zero Data Retention agreement should know that Skills are the exception — ask what that actually means in practice for anything built into a Skill.

Ask what advance notice, if any, Anthropic gives before a model changes, is deprecated, or is suddenly restricted, so the group isn't caught mid-task by a surprise.

Ask what "Claude for Science" actually is, how it differs from using Claude normally, and for plain guidance on which model fits which kind of task the group does.

Ask what happens in the failure case, not just the happy path — could a broken or expiring connection leak data or leave access in a half-open state?

Before linking any outside system to Claude, ask exactly what data becomes visible once the connection is live, not just what the connection is nominally "for."

Ask how to steer Claude toward trusted, preferred sources rather than whatever it already knows — and note that a well-maintained context graph, like this vault itself, is one working answer to exactly this question.

If any working-group member is in the UK or EU, ask specifically how GDPR and data-residency rules change the retention and training picture compared to what's documented for U.S. users.

Ask how a rare disease's information actually gets built into what a model already knows, how to regularly test whether the model answers likely patient and clinician questions well, and how that knowledge gets updated over time.

Ask Anthropic to explain the "LLM Wiki" concept the group has heard of, and how guardrails would actually be implemented around it in practice.

Ask how to keep each member organization's own sensitive files walled off from the others, while still letting the group collaborate on shared learnings, templates, and best practices.

Ask directly whether Anthropic offers reduced pricing, credits, or a grant program aimed at nonprofits or research organizations, since budget is a real constraint for this group.

Ask whether Anthropic offers any structured onboarding or training aimed at staff who aren't technical by background, rather than generic developer-facing documentation.

Not everything sensitive is PHI — ask what protections exist, or don't, for other confidential material like unpublished manuscripts, embargoed data, or funder communications.

Ask whether Claude can help design branching survey logic — skip patterns, conditional questions — in a format ready to drop into the survey tool the group already uses.

Ask for practical guidance on calibrating trust in an answer — how to tell when the model is confident versus guessing, and what assumptions might be baked into a given response.