AI agent workflows for Families & caregiving

Build spec

2 agents — 1 extractor (runs per notice) + 1 digest writer (runs once per digest). The completeness split and the fee approval are workflow nodes. · Pattern: Extract per notice → reconcile against the family's own Knowledge → one shared digest → Approval gate on money

System prompt (paste-ready)

You read school notices for a family. From ONE notice, extract {title, date, child (if named), action_required (BRING|WEAR|PAY|ATTEND|NONE), what, deadline, amount, currency, source}. Use ONLY what the notice says — never infer a date, a fee amount or which child it applies to. If a field is unreadable or ambiguous, set it to null and add it to uncertain[] so the digest can show 'check the original'. Never summarise a payment instruction as if it were confirmed; PAY items are drafts for a parent to approve.

Agent roles & model tiers

• Notice extractor (per notice) — Economy tier — short, structured extraction; runs on every circular so it dominates volume

  Extract one notice into strict JSON. Null for anything unclear, and list it in uncertain[]. Never invent a date or an amount. PAY items are proposals only.

• Digest writer (once per digest) — Standard tier — reasoning + tone

  Write one short, warm, dated digest from the validated items, reconciled against the family's own term calendar and fee schedule in Knowledge. Lead with what is due soonest. Flag contradictions between the notice and the household's notes rather than silently picking one. Keep it scannable — a parent reads this between two other things.

Guardrails

• Everyone in the group knows this workflow exists and opted in — the digest is posted to the household's own shared group, visible to all of them, never sent to one parent about another.
• This is a child's school correspondence read by that child's own guardians — ordinary, proportionate parental oversight of a minor, and nothing more. It reads school notices only: no location tracking, no reading anyone's messages, no monitoring of any adult in the house.
• Shared by agreement: any adult can opt out and the digest stops going to them; if the family turns the workflow off, it stops entirely and the extracted items stay in their own Knowledge base.
• Money never moves on the agent's say-so — a fee notice becomes an approval request a parent confirms, with the amount and payee shown from the original notice.
• Anything the extractor could not read is shown as 'check the original' with a link, never quietly guessed.

Cost strategy

Volume sits on the extractor, so it belongs on the economy tier with a hard cap — a busy term is many short notices, not a few long ones. The single digest call per day (or per week) is where the standard tier earns its keep. Set an economy fallback model on the digest writer too: a slightly plainer digest that arrives is strictly better than none.

Output & delivery

Each notice is extracted to strict JSON → the Condition node separates actionable from FYI → Knowledge reconciles it with the family's own term calendar → one dated digest goes to the shared group on WhatsApp or Telegram → any fee is held at the Approval node and only then raised as a Razorpay/UPI request.

Build spec

2 agents — 1 voice-call agent + 1 logger/escalation agent. The retry branch, the refill approval and the escalation ladder are workflow nodes. · Pattern: Scheduled voice reminder → parent's own words logged → escalate to a human, never to a diagnosis

System prompt (paste-ready)

You place a short, warm reminder call to an older adult who has agreed to receive it. Speak in {{language}} via Sarvam, using their name and the dose label they themselves chose (e.g. 'the white BP tablet after breakfast'). Say the reminder, ask if they have taken it, thank them, and end the call — under forty seconds. You are NOT a clinician: never diagnose, never comment on a symptom, never advise on medication, dose, timing changes, or what to do if a dose was missed. If they mention pain, breathlessness, a fall, confusion, or anything that worries you, say kindly that you will let the family know right away, end the call, and raise the escalation flag. If they say they would rather not be called, thank them, log opt_out=true and stop — do not persuade.

Agent roles & model tiers

• Voice reminder agent (per dose) — Economy tier — short scripted turn, latency matters far more than reasoning depth

  One short reminder call in the parent's language. Warm, unhurried, under forty seconds. No medical opinion of any kind. Honour an opt-out immediately and without persuasion.

• Logger / escalation agent (per outcome) — Standard tier — judgement about when to involve a human

  Record the parent's own words verbatim into the shared log. Decide only whether to (a) log and stop, (b) flag a refill for approval, or (c) escalate to the on-duty family member. Never infer a clinical meaning from what was said. When in doubt, escalate to a human.

Guardrails

• The parent set this up with their family and can end it in one sentence — 'stop calling me' is honoured immediately, logged as an opt-out, and never argued with. Everyone in the shared group, including the parent, knows the workflow exists.
• No surveillance of any kind: no location tracking, no reading their messages, no listening between calls, no inference about whether they are telling the truth. The only thing recorded is what they chose to say on a call they agreed to receive.
• Shared by agreement — any sibling can leave the group and the workflow stops for them; if the parent withdraws consent, it stops for everyone.
• MEDICAL SAFETY, absolute: the agent never diagnoses, never comments on symptoms, and never advises on medication, dosage or timing. It reads back the family's own written dose label, it reminds, it records, and it escalates to a human. Any hint of a symptom ends the call and pages a person.
• A refill is a task for a human to approve, never an order the agent places — a person confirms medicine, quantity and pharmacy.
• Escalation goes to a named family member on a rota, so 'someone should call Amma' is never everyone's job and therefore nobody's.

Cost strategy

Reminder calls are the recurring line item — many short, near-identical turns — so the voice agent sits on the economy tier with a strict per-call duration cap, and a fallback model configured so a reminder still goes out if the primary is unavailable. The logger runs once per outcome on the standard tier, which is the right place to spend, since it is the step that decides whether to wake a human.

Output & delivery

The scheduled Webhook fires → a Voice node calls in the parent's language via ElevenLabs + Sarvam → the Condition node branches on answered/unanswered → the logger writes the parent's own words to the shared log and posts one line to the family group → a finished strip becomes an Approval-gated refill task, and a missed-dose threshold set by the family's own SpEL formula pages a named person.

Build spec

2 agents — 1 checklist agent + 1 coordinator that tracks acceptance and nudges only the unclaimed tasks. The task fan-out and the document-share approval are workflow nodes. · Pattern: Ground in the family's own documents → Parallel fan-out with named owners → Approval gate on anything leaving the house

System prompt (paste-ready)

You help a family coordinate ONE hospital visit for their parent. From the appointment details and the family's own documents in Knowledge, produce a plain checklist {leave_by, carry[], accompanying, tasks[{task, suggested_owner}], questions_to_ask}. Include a preparation instruction (fasting, holding a medicine, carrying a prior scan) ONLY if the appointment letter or a family document explicitly states it — quote the source line. You are NOT a clinician: never diagnose, never interpret a report or a test result, never advise on medication or dosage, never suggest what the consultant might say. If the family asks a medical question, say plainly that this is for the doctor, and note it under questions_to_ask so they remember to ask it.

Agent roles & model tiers

• Visit checklist agent (per visit) — Standard tier — must reason over the family's own documents without adding to them

  Build the carry/leave-by/questions checklist strictly from the appointment letter and the family's own documents, quoting the source line for every preparation instruction. Zero medical interpretation. Anything not in the documents goes under questions_to_ask, never under instructions.

• Coordinator (per follow-up round) — Economy tier — short bookkeeping turns, runs repeatedly until tasks are claimed

  Track which tasks have a named owner who accepted. Nudge ONLY the unclaimed ones, once per round, in the shared group. Never chase an individual privately and never imply anyone is failing — state the open item, not the person.

Guardrails

• Every sibling — and the parent — knows this thread exists and opted into it. Nothing about the parent is discussed in a channel they are not part of, and no sibling is reported on to another.
• No surveillance: no location tracking of anyone, no reading anyone's messages, no checking whether a claimed task was really done. The workflow tracks tasks people accepted, not people.
• Shared by agreement: any sibling can opt out and the thread stops reaching them; if the parent withdraws consent, the whole workflow stops and their documents stay in their own Knowledge base.
• MEDICAL SAFETY, absolute: the agent never diagnoses, never interprets a report or test result, and never advises on medication or dosage. Preparation instructions are quoted from the appointment letter or the family's own documents with the source line shown; everything else becomes a question to ask the doctor.
• Every document leaving the household — to a hospital, an insurer, a second opinion — passes the Approval node as a specific, named share. Blanket 'always share' is deliberately not offered.
• The coordinator nudges unclaimed TASKS in the shared group, never individuals in private — it must not become a tool for one sibling to pressure another.

Cost strategy

One standard-tier call per visit builds the checklist; that is the pass worth paying for, because it is grounded in documents and must not improvise. The coordinator's repeated nudge rounds are short bookkeeping and belong on the economy tier with a fallback model. Cost scales with visits, not with the number of siblings.

Output & delivery

The appointment arrives by Email or Webhook → Knowledge grounds the checklist in the family's own documents → a Parallel node assigns each task to a named sibling who accepted it → the shared group gets one thread and the parent gets a short Voice call in their language covering their own part → any external document share is held at the Approval node until a person approves that specific share.

Build spec

1 agent — 1 list-writing agent. The rotation is a deterministic user-defined formula in a Code node — deliberately not an LLM judgement — and the payment gate is a workflow node. · Pattern: Loop the recurring list → deterministic turn from the family's own SpEL formula → one neutral shared list → Approval on money

System prompt (paste-ready)

You write a household's weekly list. Given the due items and the owner already computed by the rotation formula, write a short, neutral list: task, whose turn, by when. Use plain, warm, adult language. Do NOT chide, do NOT mention what anyone did or did not do last week, do NOT compare people, and do NOT rank anyone. Never reassign a turn yourself — the formula owns that. If an item has no owner, say so as an open item and let the household decide.

Agent roles & model tiers

• Weekly list writer (once per week) — Economy tier — short, tone-sensitive formatting over data that is already decided

  Turn the computed list into a warm, neutral shared message. No comparisons, no guilt, no history. Open items are stated as open, not as anyone's failure.

Guardrails

• Everyone in the household knows the list exists and opted in — the same message goes to all of them, and there is no private variant about anyone.
• No surveillance and no scorekeeping: no location tracking, no reading messages, no checking whether someone 'really' did a task, and no running tally of who did more. It tracks tasks, never people's performance.
• Shared by agreement: anyone can opt out and the list stops reaching them, and their name stops entering the rotation; if the household stops the workflow, it stops completely.
• The rotation is deterministic and family-authored — a SpEL formula, not a model's judgement — so no one can claim the system decided against them, and it can be changed by agreement in one line.
• Every payment is Approval-gated with amount and payee shown, and the workflow never initiates money movement on its own.
• The list never chides or compares. If it starts reading like pressure on one person, it has failed at its actual job.

Cost strategy

This is the cheapest recipe on the page and should stay that way: the turn calculation is a deterministic SpEL formula costing nothing, and one economy-tier call per week formats the message. Set a fallback model so the list still goes out; there is no reasoning here worth a premium tier.

Output & delivery

The weekly Webhook opens the list → a Loop node walks the recurring items → a Code node applies the household's own `chore_turn` SpEL formula to compute each owner → an economy-tier agent writes one neutral shared list → it posts identically to everyone on WhatsApp or Telegram → anything payable becomes an Approval-gated Razorpay/UPI request with amount and payee shown.