Common foundation
Setting safeguards
Six rules written once and for all, applied across all your uses: never overwrite, never delete, never send, never write without validation.
Screenshot of a real tool · fictitious data, names blurred
The problem
An academic hospital physician handles co-authors’ manuscripts, student lists, department schedules and sometimes health data. An AI that overwrites a version, sends an email or changes a calendar without validation can destroy in one second the trust built up over a month.
Figure. The AI prepares, you validate; your decisions become rules.
Step-by-step method
- Write the safeguards into your Claude preferences (Settings → Profile), so that they apply to every conversation. Example baseline:
- never delete a file: move it to
90_ARCHIVE; - never overwrite a file: any corrected version becomes
<name>_INITIALS-YYYYMMDD.<ext>next to the original; - mailboxes are read-only: no sending, no drafts, no labels; replies are proposed as text to copy;
- calendar: writing only after an explicit “accept”;
- institutional tools (department schedule, patient record): read-only;
- scheduled tasks: never created or modified without approval, and with a maximum budget.
- never delete a file: move it to
- Copy them into a file
_SYSTEME/00_REGLES.md, which every skill and every scheduled task reads first. - Require epistemic honesty: distinguish “verified / estimated / probable”, leave a “to be provided” box rather than invent, cite the DOI and PMID for every reference.
- State the real status of every action, one of five: to prepare, prepared, executed, verified, blocked. A failure is reported as it is (“executed, verification failed”). Four confusions to rule out:
- referenced ≠ opened ≠ analysed: an image only seen in a folder listing is never described;
- text generated ≠ file saved;
- link built ≠ link tested;
- deployed ≠ tested in operation.
- Make the search traceable before concluding: each relevant source gets a status (queried with a result, queried without a result, inaccessible, not queried, with the reason). “Cannot be found” is allowed only once the whole reachable scope has been covered; otherwise write “not found in the sources consulted”.
- Open and close cleanly: to resume a long thread, three lines first (inputs available, latest version kept, what is missing); at the end of the work, a report under four headings: Result · Location · Verification · Caveats. Without a “Verification” line, the deliverable has not been delivered.
- Show rather than ask: before producing anything, Claude looks for an example you have already approved, follows it and says which one it used. When in doubt, it asks for an example rather than sending a questionnaire; if questions are needed, one to three, grouped, each with a default answer. In a scheduled task, no questions: the assumption is written at the top of the report. → Framing a dictated request
- Keep secrets out of the system: no username, password, tax number or other secret in a skill, a memory or a dashboard. Such data are read from a local sheet, and a password manager is better than a plain-text file.
- Keep your hats separate: anything that belongs to industry-paid activity does not go into academic reports; attribution follows the payer, not the tool.
Deliverable
A block of preferences and a rules file, short, reviewed and dated.
Health data
No identifying patient data may be sent to a service your institution has not authorised. Pseudonymise locally, handle only aggregates in the AI, and check your institution’s policy with its data protection officer (DPO).
Usage levels, reference framework (GDPR, HDS, AI Act) and agent security: Security & confidentiality.
Starter prompt
Review my preferences and list every action you could take without my explicit approval. For each one, propose a one-line safeguard rule.