Casserole Command

Ask first

• Which children, schools and teams? (default every sender with a school or league domain)
• What counts as urgent? (default due within 48 hours)
• Which calendar? (default the shared family calendar)
• Should forms be saved somewhere? (default a folder per child)

Needs

• School and team email in one mailbox. A forwarding rule from a second parent's inbox is enough.
• A family calendar it can write to and a way to text you; a notification if it cannot text.
• A list of message ids already handled, so nothing is entered twice.
• Connecting email: "It reads only school and team senders and never replies to them."
• If your assistant can only chat: paste each school email and ask "what dates and forms are in this, and for which kid?"

Build

1. Read: on each new email from a listed sender, and in a sweep at 5pm for anything missed, fetch the message and attachments.
2. Decide: the model returns child, action, due date, and item to bring, or NONE. Keep "permission slip due Friday"; drop "great job at the bake sale". Use the grade or teacher name to pick the child; ask in the 5pm text when it cannot tell.
3. Decide (undated): an ask with no date becomes a "date pending" entry on the next Monday and is re-checked daily until a date appears.
4. Deliver: one calendar entry on the due date with a reminder at 7pm the evening before, and the attachment saved to that child's folder. Text only for items due within 48 hours.
5. Memory: store the message id and entry id; a corrected email from the school moves the entry rather than adding a second one.

Check

• Forward yourself a made-up school email with a date and run with delivery off: one entry for the right child, a text queued only if it falls within 48 hours.
• A newsletter with no asks produces nothing.
• Quiet run: an evening with no school mail writes nothing and sends nothing; the handled-id list shows the sweep time.

In my version

• Three schools (high, elementary, and a middle school next year) and two coaches. My husband forwards his school mail to mine. Family calendar, texts to me.

Ask first

• Who is the log for, and have they agreed to it in writing in the chat? (default no build without a yes)
• What does a day's entry hold? (default a 0-10 pain score and an optional note)
• How many past flares can you date? (default at least two; fewer means a simple 3-day rising rule)
• What should change when they say OK? (default their driving and cooking slots for 5 days)

Needs

• A daily score the person sends with one tap or one text; a missed day counts as unknown, never as zero.
• Sleep hours from a wearable or a self-reported number; without it the rule uses scores alone.
• Dates of past flares, from memory or old messages.
• A coding agent to build a tiny log (one table: date, score, note, sleep) and a scheduled job.
• Permission: "It reads only what he sends to the log, never his medical records, and texts only the two of us."
• If your assistant can only chat: he texts a score each evening into one thread, and once a week you ask "does this look like the week before a flare?"

Build

1. Build the log: a text number or shortcut that stores date, score and note. Reply "logged" so he knows it took.
2. Read: at 9pm, the last 7 days of scores, sleep and notes, plus the 10 days before each past flare.
3. Decide (match): fire when at least 3 of the last 4 days have a score 2 points above his 30-day median, or when the 4-day pattern is closer to a pre-flare window than to any normal window. Unknown days never count toward firing.
4. Decide (cool-off): no second text within 5 days of the last one unless the score reaches 8.
5. Deliver: one text to both, with the numbers, the past flare it resembles, and the proposed swap. Nothing changes until he replies OK.
6. Close: on OK, edit the driving plan and meal list for 5 days and text what moved. On no, log it as a false alarm; three false alarms in 60 days raises the threshold by 1.

Check

• Load last year's scores with delivery off: it fires within 4 days before each known flare and no more than twice in quiet months.
• Send a test score with a note: the log replies "logged".
• Quiet run: a normal night sends nothing; the log shows the 9pm read time.

In my version

• My husband has Crohn's disease and chose the 0-10 score himself. A coding agent built the log and job on my home server. Driving plan and meal list are shared notes the agent edits after his OK.

Ask first

• Which appliance is the parent sure to use every morning? (default the kettle or coffee maker)
• By what time? (default 10am; 11am on weekends)
• Who checks in person, and have they agreed? (default one neighbor or relative who said yes)
• How long can you cancel before they are asked? (default 30 minutes)

Needs

• A smart plug with a readable power log on that appliance. Without one: a daily "I'm up" text from the parent, with the same rule.
• The parent's calendar, so mornings out are not alarms.
• A coding agent to poll the plug every 5 minutes and keep a small log.
• Permission: "It reads one plug's power use and the parent's texts to you; it texts the neighbor only on your timer, using words you approved."
• If your assistant can only chat: there is no unattended version; agree a morning text with the parent and ask the assistant each day at 10am to remind you if it hasn't come.

Build

1. Read: poll the plug every 5 minutes from 6am. A boil is power above 1 kW for at least 60 seconds.
2. Decide: at the deadline, no boil, no text from the parent since midnight, and no calendar event before noon means fire. A calendar event or travel day skips the morning.
3. Deliver (you): one text with what is missing and the time the neighbor will be asked.
4. Decide (wait): any boil or parent text in the wait closes it. Your reply WAIT adds 60 minutes; OK closes it.
5. Deliver (neighbor): the approved text you wrote once at setup, never new wording. Then text you that it went.
6. Memory: one row per morning with first-boil time, so drift (later every week) shows in a monthly line.

Check

• Point it at a past day's log with sending off: a morning with a boil at 7:40 closes at 7:40.
• Unplug a test plug and set the deadline to 10 minutes from now: you get the text, and the neighbor draft shows as queued, not sent.
• Quiet run: a normal morning sends nothing and logs one boil time.

In my version

• Mom is 81 and lives 400 miles away; she chose the kettle. Plug polled from a small home server, Carol next door agreed, 30-minute window. In 14 months it has asked Carol twice.

Ask first

• Whose medications? (default everyone you manage, with their consent)
• How early should it warn? (default 21 days before an approval expires, 7 days before supply runs out)
• When can you make calls? (default a 30-minute hold on the next weekday morning)

Needs

• A medication list: name, who takes it, supply days, refill or approval end date. Build it once from pharmacy emails and labels; ask for anything missing.
• Pharmacy and insurer emails in one inbox, forwarded if they go to someone else.
• Connecting email: "It reads pharmacy and insurer mail only and never replies or calls."
• If your assistant can only chat: keep the list in the chat and ask each Monday "what runs out or expires in the next three weeks?"

Build

1. Read: on each new email from a pharmacy or insurer, and at noon daily, load the list and any message since the last run.
2. Decide (status): mark each item ok, waiting, stalled, or denied. Stalled means a request older than 5 business days with no reply, or a refill not shipped 4 days before supply ends.
3. Decide (urgent date): for each stalled or expiring item, the latest day a call still prevents a gap, counting 7 business days for approvals.
4. Deliver: a text with the person, item, urgent date and a three-line phone script, and a calendar hold at the next free weekday morning. Denials go to the appeal routine instead.
5. Close: a shipped or approved email marks the item ok and removes the hold. If still stalled 3 days after the text, one more text; never a third.

Check

• Add a fake item expiring in 10 days and run with delivery off: a text and hold are drafted for the right date.
• Paste a shipped email: the item flips to ok.
• Quiet run: a day with everything ok sends nothing; the list shows the noon check time.

In my version

• My husband's infusion approval and two of Mom's prescriptions, from three pharmacies. Holds go on my work calendar at 8:30am.

Ask first

• Has the parent agreed, and can you get view-only access or alert emails? (default alert emails sent to a shared inbox, no login)
• What is unusual? (default a new payee, gift cards, wires, crypto, or a charge 3 times the payee's median)
• Who else should see it? (default you only; a sibling note weekly)

Needs

• The bank's transaction alerts as emails, or a view-only statement export. Many banks let a trusted contact receive alerts; with only monthly statements, run it monthly.
• A list of the last 12 months of payees and amounts, built from the first statement.
• Permission: "It reads alerts for one account and never logs in, moves money or contacts the bank or the parent."
• If your assistant can only chat: paste the month's transactions and ask "what here does not fit her usual pattern?"

Build

1. Read: at 7pm, every alert since the last run.
2. Decide (flags): new payee; any gift card retailer or prepaid card wording; wires, transfers to a person, or crypto; any charge 3 times that payee's median; several round amounts at one place within an hour.
3. Decide (context): a calendar event or a note from you ("new plumber Tuesday") explains a new payee and drops the flag.
4. Deliver: one text listing each flag, why it fits a scam pattern, and one gentle question to ask her. Nothing goes to the parent or bank.
5. Close: FINE adds the payee to her usual list. FRAUD adds a 15-minute call hold for tomorrow 9am and a script for the bank's fraud line.
6. Memory: alert ids handled; usual payee list refreshed monthly.

Check

• Feed it a month of past alerts with sending off: zero or one flag on a normal month.
• Add a fake alert of four $100 gift card charges: one text drafted.
• Quiet run: a normal evening sends nothing; the handled ids show it ran.

In my version

• Mom said yes after a friend lost money to a phone scam. Her bank sends alerts to a shared inbox I read; my brother sees the weekly note.

Ask first

• What makes something needs-me-today? (default from your manager, a deadline within 24 hours, or a question addressed only to you)
• What is read later? (default newsletters, all-staff, and anything you are only copied on)
• When? (default 30 minutes before you start)

Needs

• Work email with permission to move messages to folders. If your employer blocks connected assistants, skip this routine.
• Your calendar, so a meeting today with the sender raises their mail.
• Permission: "It moves messages to folders and never replies, deletes or forwards."
• If your assistant can only chat: paste the subject lines and senders each morning and ask for the three piles.

Build

1. Read: every message received since 5pm yesterday.
2. Decide: sort by the rules above; when unsure, choose the higher pile.
3. Deliver: move each message and text a count with one line per needs-me item, 6 lines at most.
4. Memory: when you drag a message to a different pile, store sender and subject words as a correction rule for next time.

Check

• Run it with moving off on yesterday's mail: the needs-me list matches what you would have picked.
• Drag one message: tomorrow's dry run applies the rule.
• Quiet run: no overnight mail sends "Nothing new overnight."

In my version

• A payroll manager's inbox at a regional utility. My company allows a read-and-file assistant; nothing is sent.

Ask first

• What is the floor price per item? (default 85 percent of asking)
• When can buyers pick up? (default your free weekend slots)
• Which phrases mean a likely no-show? (default asks to hold without a time, or a shipping offer)

Needs

• Buyer messages in one inbox, from the marketplaces you sell on.
• A sheet: item, asking, floor, status, photos.
• Your calendar for pickup slots.
• Permission: "It drafts replies and never sends or changes a listing."
• If your assistant can only chat: paste each buyer message with the item line and ask for a reply.

Build

1. Read: every 2 hours, new buyer messages since the last run.
2. Decide: match each to an item. Price at or above asking: text you now. Between floor and asking: draft a counter halfway. Below floor: draft a polite no with the floor. Hold-without-time or shipping offers: flag as likely no-show.
3. Deliver: drafts in the inbox; nothing sends until you press send.
4. Close: when you mark an item sold on the sheet, draft a pickup message with two free slots and the address after they confirm.

Check

• Paste three test messages at different prices with drafting into a test folder: each gets the right reply.
• Quiet run: no new messages makes no drafts.

In my version

• About 6 pieces listed at once on two marketplaces; floors on my sheet; pickups on Saturday mornings.

Ask first

• Any diets in the house? (default none; ask each person)
• How many nights to plan? (default 5, leaving 2 for leftovers or takeout)
• How long can a weeknight dinner take on a busy night? (default 20 minutes)
• Order groceries or just list them? (default a draft cart you approve)

Needs

• The family calendar for late nights and away games.
• A list of meals that worked, built from what you tell it after each week.
• Any health flag you share, for example a symptom average. Without one, plan for the default diet.
• Permission: "It reads calendars and fills a cart; it never places the order."
• If your assistant can only chat: tell it your week and diets on Sunday and ask for 5 dinners and a list.

Build

1. Read: Sunday afternoon, next week's events, the symptom average, last week's notes on meals, and the pantry list.
2. Decide (nights): a night is busy when an adult or kid has an event within an hour of 6pm.
3. Decide (meals): busy nights get 20-minute meals; if the symptom average is 4 or above, every night has a low-residue version. No meal repeats from the last 2 weeks.
4. Deliver: the plan and list in the shared note; the cart drafted, not ordered.
5. Close: Friday, ask one question in the note, "which dinners worked?" and add them to the list.

Check

• Run it with ordering off: every busy night has a quick meal and nothing repeats.
• Set the symptom average to 5 in a test: every dinner shows a low-residue line.

In my version

• Five dinners for five people; my husband's low-residue list comes from his dietitian handout.

Ask first

• Who can drive? (default both adults plus any carpool families)
• Which signal says an adult should drive less? (default a symptom average of 4 or more, or a meeting past 5pm)
• How fair should the tally be with other families? (default within 2 trips a month)

Needs

• Every ride in the calendar, each needing a driver.
• A month's tally of who drove for whom.
• Any health trend the person has agreed to share; without one, use meetings only.
• Permission: "It reads calendars and drafts texts; nothing reaches another parent until you approve."
• If your assistant can only chat: list the rides and who is free and ask for a fair plan.

Build

1. Read: next week's rides, adult calendars, the symptom trend and the tally.
2. Decide: give each ride to a free adult. An adult with a high trend gets at most 2 rides. If no adult is free, choose the carpool family who owes the most trips.
3. Deliver: the plan in the shared note, and one draft per outside family for you to approve.
4. Close: a yes marks the ride and updates the tally; no reply by Tuesday noon re-drafts the ask to the next family.

Check

• Run on next week with sending off: no adult has two rides at once, and outside asks go to the family owed most.
• Set the trend high: that adult drops to 2 rides or fewer.

In my version

• Three kids, two carpool families, tally in our shared note. Re-runs when Dan says OK to the flare text.

Ask first

• Which day and time? (default Friday 8am)
• Where should picks arrive? (default where you already hear from your assistant)
• How many picks at most? (default 3)

Needs

• An assistant that can run on a schedule and open a web page.
• What your assistant knows about you: your routines, tools and recurring tasks.
• If your assistant can only chat: ask "anything new on Runbook for me?" once a week and it does the same check then.

Build

1. Read: open https://myrunbook.ai/new.json?since=<date of the last check, or 7 days ago>.
2. Decide: pick at most three routines that fit something specific about you: a task you repeat, a tool you already use, a gap in your week. Skip ones you already have and ones needing a setup you lack.
3. Deliver: if nothing fits, send nothing. Otherwise one short message: each pick's name, one line on why it fits, its link, ending "Reply with a number and I'll set it up."
4. Close: never build without a yes. On a yes, open that routine's build steps and follow them.
5. Memory: save the date of this check for next week.

Check

• Run it now with since set to 30 days ago and delivery off; every pick must name a specific reason it fits.
• Quiet run: a week with nothing fitting sends no message.

In my version

• A scheduled assistant, Fridays at 8am, picks sent by text.

Ask first

• Which children? (default those with an online grade portal)
• What counts as a drop? (default a full letter grade or 5 points in a week)
• Who gets the note? (default both parents only)

Needs

• Grade portal access per child, or the weekly emails the portal sends. Without either, teacher emails only.
• Permission: "It reads grades and never messages teachers or the kids."
• If your assistant can only chat: paste the portal page on Friday and ask for missing work and drops.

Build

1. Read: Friday after school, each child's assignments and grades, plus last Friday's snapshot.
2. Decide: list missing and late items and any grade below the drop rule.
3. Decide (question): one plain question per child with something open; none for a child with nothing open.
4. Deliver: one short note in the parents' thread.
5. Memory: save this week's snapshot to compare next Friday.

Check

• Run it now with delivery off against this week's portal: each missing item matches what the portal shows.
• Quiet run: a week with nothing missing sends "All three are clear."

In my version

• Three portals, two parents in the thread.

Ask first

• Has the parent named you as a proxy on their patient portal? (default no build without it; use their visit papers instead)
• How early should the brief arrive? (default 3 days before, a reminder the day before)
• Who else sees it? (default siblings who help)

Needs

• Visit summaries from the portal, or photos of the paper ones.
• A medication list and any log the parent keeps, like weight or blood pressure.
• Notes from your calls with them; a voice memo after each call is enough.
• Permission: "It reads the portal as a proxy and never messages the doctor's office."
• If your assistant can only chat: paste the last visit summary and your notes and ask for three questions.

Build

1. Read: Sundays, find visits in the next 10 days. For each, load the last summary for that doctor, medication changes since, and your call notes.
2. Decide (changes): list new or changed medicines, new symptoms mentioned twice or more, and log values moving more than the doctor's stated range.
3. Decide (questions): three questions, each tied to one change, in plain words the parent could ask.
4. Deliver: the brief on the shared page 3 days before, and a text the day before with the time and driver.
5. Close: after the visit, you send a voice note; it files a dated entry in the timeline and updates the medication list. Siblings get one line.

Check

• Run on a past visit with delivery off: every question ties to a real change.
• Send a test voice note: a dated timeline entry appears, nothing is sent to the clinic.

In my version

• Mom, 81, sees four doctors. Proxy access set up with her; my brother and sister read the folder.

Ask first

• Who gets it? (default the siblings or relatives who help)
• How many asks per person? (default one each)
• Tone? (default short and plain)

Needs

• The parent's timeline, calendar and any logs from other routines; without them, your notes for the week.
• A shared task list.
• Permission: "It drafts only; you send."
• If your assistant can only chat: paste your week's notes and ask for a short update with one ask each.

Build

1. Read: Sunday evening, the week's entries.
2. Decide: lead with one plain line on how the week went, then next week's appointments, then one ask per sibling from open tasks, rotating so nobody gets the same kind twice in a row.
3. Deliver: a draft in the chat for you to send.
4. Memory: mark each ask given; next week it checks whether each was done and mentions it.

Check

• Run it now with delivery off: one ask each, and no task repeats from last week.
• Quiet run: a week with no entries drafts "Quiet week, nothing new."

In my version

• Two siblings in other states; asks rotate from a shared list.

Ask first

• How many sessions and how long? (default 3 sessions, 40 minutes)
• Which windows are allowed? (default before 7am, lunch, or weekend mornings)
• Outdoors when? (default dry and above 35F)

Needs

• Your calendar and any shared caregiving plan.
• A forecast for outdoor sessions.
• Permission: "It adds holds to your calendar and never moves other people's events."
• If your assistant can only chat: give it your week and ask for three gaps.

Build

1. Read: next week's calendar, the driving plan and the forecast.
2. Decide: pick three gaps at least one day apart; outdoors only in dry windows; skip days with two or more rides.
3. Deliver: holds marked busy, and a text listing them.
4. Memory: count sessions you mark done; two missed weeks in a row drops the target to 2 and says so.

Check

• Run with booking off: three gaps, none overlapping a ride.
• Quiet run: a week with no room books nothing and says which days were full.

In my version

• Three 40-minute holds; runs outside only above 35F.

Ask first

• Who is on the plan? (default everyone on the family policy)
• Where do bills arrive? (default email plus photos of paper ones)
• How long before an unmatched bill is flagged? (default 30 days)

Needs

• Bills from providers and statements from the insurer, as PDFs or photos. A coding agent builds a small parser and a table.
• Your deductible and out-of-pocket totals for the year.
• Permission: "It reads bills and statements and drafts emails; you send them."
• If your assistant can only chat: paste one bill and its matching statement and ask "what do I owe and is this right?"

Build

1. Read: on each new bill and on the first Saturday, extract provider, date of service, billed, and patient share from every bill and statement.
2. Decide (match): pair by patient, provider and date within 3 days.
3. Decide (owe): pay when the bill equals the statement's patient share. Hold when there is no statement yet. Dispute when the bill is higher by $10 or more.
4. Deliver: a row per bill in the sheet and a dispute draft per mismatch, quoting the statement date.
5. Close: a later bill or zero balance closes the dispute; nothing heard in 30 days drafts a follow-up.

Check

• Load last year's bills with drafting off: each matches, and known overcharges are flagged.
• Quiet run: a month with no new bills leaves the sheet unchanged.

In my version

• Five people on one plan. Built by a coding agent; my husband's care produces most of the bills.

Ask first

• Which account type, and when does unused money expire? (default a health spending account, use by Dec 31)
• Which expenses qualify? (default copays, prescriptions, eye, dental and over-the-counter items your plan lists)
• Where do paper receipts go? (default a photo into one receipts folder)
• Who submits? (default you, after reading the draft)

Needs

• Receipts from email (pharmacy, clinic, copay notices) and a folder of photos. A missing receipt is listed by date and amount rather than skipped.
• The plan's claim form as a PDF, and a sheet of claims already filed this year.
• Permission: "It reads receipts and fills the form; it never logs in to the plan or submits."
• If your assistant can only chat: send the month's receipt photos and ask for a total by person and a list of anything missing.

Build

1. Read: on the 25th, every receipt dated since the last claim, plus the claims sheet.
2. Decide (eligible): keep items on the plan's list; drop duplicates with the same date, provider and amount; hold anything over $200 with no itemized receipt.
3. Decide (gaps): copay emails with no matching receipt go on a short missing list.
4. Deliver: the filled form, one PDF bundle in date order, and a draft email to the plan in your inbox. Nothing is sent.
5. Close: when a reimbursement email arrives, mark those rows paid. From October on, add the amount left to use before the deadline.
6. Memory: the claims sheet holds each receipt id so none is claimed twice.

Check

• Run it on last month with drafting off: the total equals the receipts you can count by hand, and no receipt appears twice.
• Quiet run: a month with no receipts drafts nothing and writes one line, "No new receipts."

In my version

• Our family spending account, five people, claims drafted on the 25th; I submit through the plan's upload page.

Ask first

• What share of profit to set aside? (default 30 percent; ask your tax preparer once)
• Which costs count against sales? (default pieces bought, paint, hardware and fees)
• Which day should the text come? (default the 1st)

Needs

• Payout records from each marketplace, as emails or a monthly export.
• A cost sheet you add to when you buy pieces or supplies; with no costs logged, it uses sales only and says so.
• The quarterly estimated payment dates for your country or state.
• Permission: "It reads payouts and your sheet; it never moves money or files anything."
• If your assistant can only chat: give it the month's sales and costs and ask how much to set aside.

Build

1. Read: on the 1st, last month's payouts and costs.
2. Decide: profit is payouts minus costs; the set-aside is profit times the rate, rounded to $10. A loss month sets aside nothing and carries the loss forward.
3. Deliver: one text with sales, costs, profit and the amount to move.
4. Deliver (dates): calendar reminders 10 days before each quarterly payment, with the year-to-date set-aside so the payment amount is visible.
5. Memory: a running year table for the tax return; a payout id is counted once.

Check

• Run it on last month: the profit matches your own math to the dollar.
• Quiet run: a month with no sales texts "No furniture sales last month."

In my version

• 30 percent rate, two marketplaces, set-aside moved by hand into a savings account.

Ask first

• Which jobs, and how often? (default a starter list for your house type and climate)
• Who does each job? (default you; mark ones for a contractor)
• Should storm warnings trigger an extra check? (default yes, for rain over 1 inch or frost)

Needs

• A house job list with last-done dates; the first run writes a starter list for you to edit.
• A 10-day forecast for your town.
• If your assistant can only chat: ask on the 1st, "what house jobs this month?" and paste your list.

Build

1. Read: on the 1st, the job list and forecast. Mid-month, re-read the forecast only.
2. Decide (due): jobs whose interval has passed, plus seasonal jobs before the first forecast frost or a storm.
3. Decide (timing): pair each outdoor job with a dry weekend day in the forecast.
4. Deliver: 3 to 6 jobs in the family note, past due at the top. A storm mid-month adds one line, not a new list.
5. Memory: you tick a job done; the date updates and it drops off.

Check

• Run now with delivery off: every job shown is past its interval or tied to a forecast.
• Quiet run: a month with nothing due writes "No house jobs due."

In my version

• A 1990s house, 18 jobs on the list, in our shared family note.

Ask first

• Who is watching, and for how long? (default one evening)
• Which medical details may appear? (default allergies, medicines and when to call)
• What should it never include? (default no account logins, alarm codes or door codes)

Needs

• What you tell it: names, bedtimes, allergies, medicines, numbers. A stored family details note saves retyping.
• If your assistant can only chat: this routine is chat already; keep the details note in the conversation.

Build

1. Read: what you give it and the stored note.
2. Decide: allergies, medicines and what to do if someone is unwell go first, then the evening in time order, then numbers.
3. Decide (gaps): any child with no bedtime or no allergy line gets a question back before the sheet is made.
4. Deliver: one printable page, large type, 25 lines at most.

Check

• Ask for one with fake details and a peanut allergy: the allergy is the first line.
• Leave out a bedtime: it asks before writing.

In my version

• A chat assistant, a few times a month; the sheet goes on the fridge.

Ask first

• What family details may it store? (default names, dates of birth, doctors, insurance member ID; never full record numbers)
• What should it do with fields it cannot fill? (default mark them, never guess)

Needs

• A stored details note in the conversation or a file it can read.
• A clear photo of the form, one page at a time.
• If your assistant can only chat: this routine is chat already.

Build

1. Read: the photo and the details note.
2. Decide: match each field to a stored detail by meaning ("guardian" means you or your partner). Mark unknown fields with where to find the answer.
3. Deliver: a numbered list in the form's order, one line per field, so you can copy line by line.
4. Memory: when you give a missing answer, ask whether to store it.

Check

• Send a sample form with fake details: unknown fields are marked, never invented.
• Give a missing answer: it asks before storing it.

In my version

• A chat assistant with a family details note; mostly school and sports physical forms.