These are your working files.
If you are reading this inside the toolkit repository, copy the folder to a private place first and rename the copy. Do not fill these templates inside the public repository.
Read ai-patient-toolkit/02-privacy-and-safety.md first.
- Turn on device encryption.
- Make an encrypted backup.
- Check the chosen AI provider's training and retention controls.
- Decide which single file the assistant may read first.
- Get consent from the person whose records these are.
- Add the protected location of the first record to
source-index.md. - Fill
00-current-state.mdfrom formal records. - Mark anything written from memory as
UNVERIFIED.
Put the originals in original-records/ next to these files. Let the assistant read two or three at a time, after you approve them. Nothing in there gets edited.
These files are the flat starting set. Numbered subfolders — diagnosis, treatment, genetics, decisions, side effects, research, archive — get added as records accumulate. The full structure is in ai-patient-toolkit/appendix/templates.md.
original-records/— your originals, as they came. Never edited.00-current-state.md— verified snapshot of the case.source-index.md— source IDs, protected locations, and approved working copies.care-team.md— names and contact details. Keep local.decision-note.md— one question, its evidence, and the clinician's decision. Copy this file for each new decision.side-effect-log.md— observations and advice received. Copy it per side effect when one log gets crowded.agents/research-watch.md— source-first research instructions.agents/appointment-prep.md— a short appointment agenda..gitignore— a basic guard against committing raw files.
Dated copies are enough.
If you choose git, keep it local. Add no remote. Check with:
git remote -vThe command should print nothing.
The .gitignore reduces mistakes with untracked files. Remove sensitive files before the first commit. Keep a separate encrypted backup.
- Share the smallest useful file set.
- Review direct identifiers.
- State the exact task.
- Require source IDs for every extracted fact.
- Verify the output against the originals.
- Ask a clinician about medical meaning.
Urgent symptoms go to a clinician or local emergency service. Close the AI workflow.