AI Patient Toolkit

This repository contains a folder structure, templates, and AI-assistant instructions for organizing medical records.

I built this during my own treatment for metastatic breast cancer. I use the same folder structure for my records. Version 0.2 publishes the reusable parts. I write about the background in Information Body.

Included: file-organization instructions, a source-checked health-profile template, decision and side-effect logs, appointment-prep instructions, and a research-lead workflow.

Use AI to organize records, map sources, and prepare questions. Diagnosis and emergency assessment require qualified clinicians. Discuss treatment choices with your medical team. For an emergency, contact the local emergency service.

Cloud assistants process relevant file content on their providers' systems, including files opened from a local folder. Medical profiles remain sensitive after names are removed. Read Privacy and processing before adding records.

Why plain text

A profile in plain text stays readable by any assistant and any text editor. Vendor health features keep your records inside one provider's system. Here the files stay on your disk, in a format you can move, compare, and back up. Processing still happens at the provider — see Privacy and processing. When a tool changes or shuts down, the folder is unaffected.

Quick start

1. Create a folder, for example my-health. Inside it, create original-records/ and copy every medical file you have into it: PDFs, photos of documents, screenshots, portal exports. Any order, any shape. Nothing in there gets edited.

2. Read the privacy and safety guide. Check device encryption, backups, and your provider's training and retention settings.

3. Open my-health as a local folder in Claude Code or Codex. These are coding assistants that can see files on your computer. You talk to them in plain language.

4. Paste this prompt:

   Read https://github.com/dariaminsky/ai-patient-toolkit/blob/main/AGENTS.md and follow it. My originals are in original-records/. Explain the data flow before reading any of them. Ask for consent for each batch of files. Show me the file plan first. Keep the originals unchanged. Ask me to verify every extracted date, unit, value, and source.

5. If the assistant cannot open links: press the green Code button on this page, then Download ZIP. Unzip it into my-health and point the assistant at ai-patient-toolkit/AGENTS.md instead. The starter/ folder holds the same templates for filling in by hand.

6. Review the plan. Approve one step at a time. Start with two or three documents, not the whole archive. Check every extracted fact against the original.

When a new report arrives, drop it into original-records/. Ask the assistant to update the profile as a new version. Review the changes.

Contents of v0.2

• Assistant setup instructions
• Mindset
• Data and versioning guide
• Privacy and safety guide
• Starter files:
  • Overview
  • Current-state profile
  • Source index
  • Care team
  • Decision note
  • Side-effect log
  • Research watch
  • Appointment prep
• Synthetic example:
  • Example overview
  • Example current-state profile
• Prompt templates
• Resources

Files and versions

The main file is 00-current-state.md. It records the current diagnosis, treatment, results, sources, and open questions.

Without it, each new chat starts from a blank context and answers in general terms. With it, the assistant reads the diagnosis, treatment history, and results first, and you can check which source each fact came from.

Save each meaningful update as a new version. A later review may need the state from a specific date. Git can record the changes. Keep a separate encrypted backup too.

The files use Markdown (.md): plain text with headings and lists. Text editors and AI tools can read it. You can compare versions and move the files between tools.

Reusable agent files

Agent files are saved instructions for a repeated task.

The research-watch file asks the assistant to look for papers and trial updates. It can miss sources or read them badly. Open the primary source. Take treatment-relevant findings to a clinician. Check the source even when two models agree.

The appointment-prep file asks the assistant to draft a short question list from the current profile. Review every question before the visit.

Run these files manually in v0.2. Scheduling depends on the tool.

Research monitoring is optional. Skip it if it adds work you do not want.

Privacy and processing

Choosing a local folder controls where the files are stored. A cloud assistant still sends relevant file content to its provider for processing.

Removing names lowers the risk. A medical profile can still identify someone through a rare diagnosis, genetics, or dates.

• Check the provider's retention and training policies.
• Remove names, birth dates, record numbers, and hospital names before web searches.
• Keep IDs, insurance numbers, and credentials out of the health folder.
• Keep the health folder separate from this toolkit repository. Never push it to GitHub.
• Read the full privacy and safety guide before adding a large record set.

About this project

I write about the background in Information Body.

Contributing

Corrections and additions are welcome. Open an issue or a pull request. Researchers can open an issue to correct a citation or its framing.

License

Guide text and examples: CC BY 4.0. Starter files and scripts: MIT. See LICENSE.md.