Your assistant,
set up once.

The recipe for building an assistant on your own data, in the harness you already use. Our example is health; the same structure works elsewhere.

Last updated: October 2026

See the recipe What it doesn’t decide

The short
version.

You don’t need another app. Your apps and documents become sources, gathered in one project, with instructions written once. Then you ask, and the answer takes the shape the question needs: a trend table, a summary, a list of questions.

The structure doesn’t get rebuilt. Every new piece of data adds to the history, and when a better model comes out, it reads the same files with the same instructions.

New to the word harness? Our guide “What surrounds the model” explains it.

The
structure.

Sources go into a project. The instructions tell the model how to use them, the model answers in the shape you ask for, and the history builds up month after month.

Sources go into a project, which answers in several shapes Sources watch, ring scale blood work (PDF) notes, logs Project in Claude or ChatGPT Instructions written once Model reads, compares, answers Outputs trends dashboard summary questions History: every new export joins the project’s files

The recipe,
in six steps.

We use health as the example because its sources are many and scattered. The steps are the same for another subject.

  1. 01

    Give it a place.

    A project keeps the files, instructions and conversations for one subject together. Claude and ChatGPT both have them. Claude’s free plan allows up to five projects, and the number of files per project in ChatGPT depends on the plan. For a history that grows, a paid plan soon becomes necessary.

  2. 02

    Gather your sources.

    Export what your devices measure and drop it into the project. WHOOP: menu, Data Export, pick the range; a link to a CSV file arrives by email. Withings: the Health Dashboard on the web, your profile, Settings, then Download my data; the data comes as CSV. Add your blood-work PDFs, the notes your therapists give you, and a document of your own: training, meals, supplements, history, goals.

    In January 2026, Anthropic announced beta integrations with Apple Health, Health Connect and two medical-records services for Pro and Max subscribers in the US. The announcement names no other country; elsewhere, exports do the job.

  3. 03

    Write the context once.

    The project’s instructions are read at the start of every conversation. Anything that doesn’t change from week to week goes there. A starting point, to adapt:

    You help me follow my health from the files in this project.
My goals: [your goals]
My history: [conditions, injuries, ongoing follow-ups]
Supplements: [what you take]
Meal preferences: [what you eat, what you avoid]
Answer briefly, in English, with a table when there are numbers.
For every number, name the file and the date it comes from.
You make no diagnosis and suggest no dose or treatment. When a question belongs to my doctor or my therapists, say so and help me prepare my question.
  4. 04

    Ask for what you need.

    The shape follows the question. For example: “Make a table of my sleep and recovery over the last 90 days, week by week.” Or: “What changed between my last two blood tests? List the differences without interpreting them.” Or: “Prepare my questions for Thursday’s appointment with my physio.”

  5. 05

    Keep it current.

    Add a new export whenever you want an up-to-date picture, monthly for instance, and every new blood test. When the vendor releases a better model, there’s nothing to redo: it reads the same files and the same instructions.

  6. 06

    Check before you rely on it.

    A model can misread a table or a PDF. Since the instructions ask it to name its source, open the file and compare when a number matters. Re-run a few checks when the model changes: a new model may answer differently.

What it
doesn’t decide.

An assistant set up this way reads, compares and summarizes. It replaces neither your doctor nor your therapists: a diagnosis, a dose or a change of treatment is decided with them. This guide is not medical advice.

Your health data is among the most sensitive you have. Before putting any into a project, read the vendor’s privacy policy and check your settings. In Claude, the Help Improve Claude setting, under Settings then Privacy, decides whether your new conversations are used to train models. In ChatGPT, it’s Improve the model for everyone, under Settings then Data controls.

The same structure,
elsewhere.

Swap the sources and the instructions, and the recipe holds: a household’s statements and budget, your customers’ history, the upkeep of a house. For a business, the structure is the same, with more sources, access set per person and checks before any action goes out.

For your business

Setting this structure up on a team’s systems, with the right access and the right checks, is what we do. Start with a conversation.

Discuss your challenge

Sources

Product facts were read on the vendors’ own pages on October 13, 2026. The example questions and the instructions template are ours.