Family history · Import

Build your tree from a long description

Have an older relative who keeps it all in their head? Record what they tell you, paste it into your own ChatGPT or Claude, and bring the result back here. Almavivo never sees the description — the LLM call happens in your own account, and only the structured JSON comes back to your device.

How this works

Copy the prompt below. Open ChatGPT, Claude, or any other LLM you already use.
Paste the prompt, then paste your family description underneath it where it says “FAMILY DESCRIPTION”.
Send. The model returns a JSON block.
Copy the JSON, paste it below, and click Validate & import.

We do not run an LLM ourselves and your description is never sent to Almavivo — see transparency for the network-level audit.

Step 1 · Copy this prompt

Tap Copy prompt, then open any chat model — chatgpt.com, claude.ai, or a local model running on your own computer (see below). Paste the prompt, add your family description at the bottom, and send.

Preview the prompt

You are helping me turn a spoken family-history description into a structured JSON graph for a privacy-first health tool. The current year is 2026. Read the description below "FAMILY DESCRIPTION" and output ONLY a JSON object — no prose, no markdown fences — matching this exact schema:

{
  "persons": [
    {
      "id": "p1",                                  // any short string unique within this payload
      "displayLabel": "Grandma Rose",              // name or however they were referred to (max 80 chars)
      "sexAtBirth": "male | female | unknown",
      "status": "alive | deceased | estranged | adopted_out_unknown | unknown",
      "currentAge": number or omit,                // if alive
      "ageAtDeath": number or omit,                // if deceased
      "birthYear": number or omit,
      "deathYear": number or omit,
      "causeOfDeath": "cv_heart_attack | cv_sudden_cardiac | cv_heart_failure | cv_stroke_ischemic | cv_stroke_hemorrhagic | cv_stroke_unknown | cv_aortic_thoracic | cv_aortic_abdominal | cv_pulmonary_embolism | cancer | cancer_unknown_type | endo_t2d_complications | resp_copd | resp_pneumonia | neuro_alzheimers | neuro_other_dementia | neuro_parkinsons | ext_accident | ext_suicide | natural_old_age | other_specify | unknown | omit",
      "causeOfDeathCancerType": "breast | ovarian | uterine_endometrial | cervical | prostate | colorectal | stomach_gastric | pancreatic | liver | lung | kidney_renal | bladder | thyroid | brain_cns | melanoma | leukemia | lymphoma | multiple_myeloma | sarcoma | small_bowel | urothelial | biliary | other | unknown_type | omit",  // only if causeOfDeath === "cancer"
      "smoking": "current | former | never | unknown | omit",
      "diagnoses": [
        {
          "condition": "cancer | myocardial_infarction | stroke | thoracic_aortic_aneurysm_or_dissection | abdominal_aortic_aneurysm | type_1_diabetes | type_2_diabetes | very_high_cholesterol_young | familial_hypercholesterolemia_known | early_onset_dementia | alzheimers | parkinsons | huntingtons | polycystic_kidney_disease | hereditary_haemochromatosis | cardiomyopathy | long_qt_or_arrhythmia_syndrome | advanced_adenoma_or_polyp | other",
          "cancerType": "breast | ovarian | uterine_endometrial | cervical | prostate | colorectal | stomach_gastric | pancreatic | liver | lung | kidney_renal | bladder | thyroid | brain_cns | melanoma | leukemia | lymphoma | multiple_myeloma | sarcoma | small_bowel | urothelial | biliary | other | unknown_type | omit",  // only if condition === "cancer"
          "ageAtDx": number or "unknown",
          "confidence": "medical_record | family_told | best_guess | unknown"
        }
      ],
      "medicalNotes": "clinical free-text NOT shown on the wall chart (max 400 chars) or omit",
      "keepsakeNotes": "one short sentence of family-tree colour (max 200 chars) or omit"
    }
  ],
  "partnerOf": [{ "a": "p1", "b": "p2" }],         // unordered partner pairs
  "parentOf": [{ "parent": "p1", "child": "p3" }]  // who is the parent of whom
}

Rules:
- **Include the user with id "__self__".** This is the anchor that lets the importer connect the rest of the family to the existing "this is me" row in the tool. Give "__self__" the same fields any other person gets (sexAtBirth, status, displayLabel if known, diagnoses, etc.) — the importer will copy the description onto the user's own row. partnerOf and parentOf edges that reference "__self__" attach to the user, so e.g. the user's mother is a parentOf {parent: "p_mum", child: "__self__"}.
- **Use partnerOf for marriages / long-term partnerships.** Don't double-list partners on each person — one entry per pair.
- **Use parentOf to attach children to their parents.** Emit one parentOf entry per parent-child pair. Do not classify the parent type — the user confirms which links are biological vs adoptive / step / foster after import, since the kind affects hereditary analysis and isn't safe to guess from a free-text description.
- Mention step / foster / adoptive relationships in the parentOf list anyway — the tool can model all of them. Note the social context in medicalNotes if it's relevant (e.g. "raised by Auntie Mae after age 5").
- **At most 2 biological parents per child** is enforced after the user confirms kinds; you don't need to limit your output.
- IDs in persons[].id are scratch — they only need to be consistent within this payload so partnerOf / parentOf can reference them.
- Map "first cousin once removed", "great-aunt", etc. to the closest plausible structural position and mention the original wording in medicalNotes.
- Sensitive causes of death (suicide, overdose) belong in medicalNotes, not keepsakeNotes — the wall chart is a family showpiece.
- Use null/omit, never make up ages or causes that weren't stated.
- Output strictly valid JSON. No trailing commas, no comments, no text outside the JSON object.

This is a graph-shaped schema; smaller LLMs sometimes mangle it. If the result fails to import, paste the prompt into a larger model (Claude Sonnet/Opus, GPT-4-class).

FAMILY DESCRIPTION:
[paste the long description here, then send to the model]

Don’t want to share your family with ChatGPT or Claude? Run a local model instead.

The prompt is plain text and works with any chat model. To keep relatives’ names off commercial services, run a model on your own machine with Ollama. The new graph-shaped schema is harder for small models, so prefer the larger options — gemma3:12b or qwen2.5:14b. If a small model struggles, paste the same prompt into Claude or ChatGPT.

Step 2 · Paste the JSON the model returns

A note on accuracy

LLMs are good at structuring text but can mishear “great-aunt on my father’s side” vs “aunt on my father’s side”, or assign a generic cause when the description is fuzzy. After importing, scan the entries on the family history page and correct anything that looks off — the pedigree chart and pattern detection are only as good as the inputs. Confidence on each diagnosis defaults to “family_told” for that reason.