LAWebsite

Ask first

• Which team app, and for which child? (default the one app your kid's team uses)
• Whose attendance do you manage? (default your child, yourself, and a partner)
• What counts as a conflict? (default any overlapping family-calendar event: trips, appointments, other activities)
• Whose travel counts, and how does it show? (default a partner whose trips appear as calendar entries naming a place)
• Fix quietly or tell you each time? (default quietly, no play-by-play)

Needs

• The family calendar. Connecting it lets the assistant read event titles, times and places; it does not change events.
• A team-app login that can set RSVPs. Fallback: the assistant sends you the list and you tap the changes yourself.
• If your assistant can only chat: paste the week's team events and your calendar, ask which RSVPs should change, and tap them yourself.

Build

1. Read: each morning, read the team app's events for the next 30 days and the family calendar for the same window.
2. Decide: an event conflicts when a family-calendar entry overlaps it. All-day or multi-day entries naming a place mean the people on them are away those days.
3. Decide: set each managed person to No when a conflict or their own travel covers the event; leave them Yes otherwise. Past events and RSVPs already correct stay untouched.
4. Deliver: apply the changes in the team app. Never message the coach or other parents, and don't report conflicts; the fix is the whole job.
5. Read: also run when a new travel entry appears on the calendar, not only at the morning check.
6. Memory: note the last applied state per event; only change what differs.

Check

• Put a fake all-day "trip" over the next practice and run: that practice flips to No for the people away. Delete it and run again: they flip back.
• A normal week changes nothing.

In my version

• Instinct, the assistant I text, with my calendar connected and my saved TeamSnap login (the app my daughter's team uses). I manage three people: her, my husband and me. His trips show on my calendar as entries naming a city, so those days he is No. It fixes everything quietly; I asked it to stop telling me about conflicts.

Ask first

• Which library accounts? (default every card your household uses)
• When is a book urgent? (default when renewals are used up or another reader's hold blocks renewal)
• How far ahead do you want the warning? (default a week before the due date, then once more just before)
• Where should it land? (default a text)

Needs

• A way to read each account's checkouts, due dates and renewals left. Emailed due-date notices are the fallback when an account can't be read directly.
• Connecting an account: it reads checkouts, due dates and renewals; it never renews or places holds unless you ask.
• If your assistant can only chat: paste your checkout list and ask what must go back.

Build

1. Read: on each check, read every checkout's due date and renewals left across all accounts.
2. Decide: a book must go back when renewals are used up or a hold blocks renewal. That is the whole rule; a book that can still renew produces nothing.
3. Decide: warn a week before the due date of a must-go-back book, then once more the day before. Books due together go in one message.
4. Deliver: a text listing the books, the due date and the library. No message when nothing qualifies.
5. Memory: note what you already warned about; warn again only on a new due date or a newly must-go-back book.

Check

• In a copy of the checkout list, set one book to zero renewals left, due in 7 days, and run with texting off: the output lists that book only.
• A list where everything can still renew produces no warning.

In my version

• Instinct with saved logins for our two public library systems, reading the live accounts; emailed due notices are the fallback. It renews nothing and places no holds; it only tells me what has to physically go back.

Ask first

• What time should it arrive? (default 7:30am)
• Which calendars count? (default your personal and family calendars)
• Include weather? (default yes, in your local units)
• Include to-dos due today? (default yes, max 3)
• How should it reach you? (default text; fallback push notification or email)

Needs

• Your calendar. Fallback: paste tomorrow's plan the night before.
• A to-do list, if you keep one. If not, drop that line.
• Connecting a calendar lets the assistant read event titles, times and places; it does not change or delete events.
• If your assistant can only chat: send "debrief me" each morning with your day pasted in, and it returns the same summary.

Build

1. Read: at the set time, read today's events (midnight to midnight), the weather for your city, and to-dos due today.
2. Decide: list events in time order, max 6. Add one short prep note to any event that needs something brought, signed or paid. Flag any two events that overlap.
3. Decide: pick at most 3 to-dos, oldest due date first. If there are no events and no to-dos, the whole message is "Clear day."
4. Deliver: one text under 600 characters, to you only.
5. Memory: none needed; each day stands alone.
6. Cost: about 30 model calls a month.

Check

• Run it now for tomorrow with delivery off. Every event in the output must match one on the calendar, in the same order.
• An empty day returns exactly "Clear day."
• Confirm the first real run by checking for the text at the set time.

In my version

• Instinct, a texting AI agent, with my accounts connected. Arrives at 7:30am by text.

Ask first

• Which mailbox? (default your personal one)
• What never gets a draft? (default newsletters, receipts, marketing, and sensitive mail: money, medical, legal)
• How do you spot that a draft is waiting? (default the original is marked unread)
• Any senders who always or never get drafts? (default none; judge each email)
• A nudge for urgent ones? (default no nudges; the draft is the only trace)

Needs

• Your email with read and draft access. It reads incoming mail and creates drafts; it never sends anything.
• A few sent emails of yours, for voice. Fallback: paste three you liked.
• If your assistant can only chat: forward the email and ask for a reply in your voice, then paste it into your mail app yourself.

Build

1. Read: when new mail arrives, read it and the thread it sits on.
2. Decide: it needs a reply when a real person asks you something, invites you, or sends something that calls for an answer. Keep: a parent asking whether your kid is coming Saturday. Drop: a school newsletter that happens to mention an RSVP.
3. Decide: check your sent mail on the thread; if you already answered, skip.
4. Deliver: create a reply draft on the same thread, short and in your voice. Mark the original unread so you still see it. Never send. No texts or notifications about it.
5. Memory: note drafted threads; a new message on a drafted thread updates the existing draft instead of adding a second.

Check

• Send yourself a test email asking a question; the next run must leave a draft on that thread with the original unread.
• Send yourself a receipt; nothing appears.

In my version

• Instinct with my Gmail connected. It skips newsletters, receipts and sensitive mail, marks originals unread, and stays quiet; I find the drafts when I open my inbox.

Ask first

• Where do your to-dos live? (default the one list you actually use)
• When should the check run? (default early morning)
• What counts as schedulable? (default anything with a due date coming up, waiting on someone, or needing a block of free time)
• How many suggestions per morning? (default one, with a one-line reason)
• Can it write to the list? (default read-only; it adds or checks off items only when you tell it to)

Needs

• Read access to the to-do list. Fallback: paste the list the night before.
• Your calendar, to see open time. Connecting it lets the assistant read events; it does not add any.
• If your assistant can only chat: paste the list in the morning and ask "what needs a slot?"; schedule the answer yourself.

Build

1. Read: each morning, read the open items on the list and today's and this week's calendar.
2. Decide: flag an item when its due date is close, someone else is waiting on it, or it needs a block of free time. Keep: "book the dentist" when Thursday afternoon is open. Drop: "someday: learn Italian".
3. Decide: pick at most one item, the most pressing, with one line on why.
4. Deliver: a text naming the item and the proposed slot or action. No text on mornings when nothing qualifies.
5. Memory: note what you suggested and when; don't re-suggest an item the person declined or scheduled.

Check

• Add a test item "call the bank, due Friday" and run with texting off: the output names that item and a real open slot.
• Run again the same morning: no repeat suggestion for the same item.

In my version

• Instinct reads my personal to-do site each morning, read-only, against my calendar, and texts at most one suggestion with a reason. It writes to the site only when I tell it to.

Ask first

• Which classes, days and times? (default the 2-3 you go to most)
• When does registration open? (default 7 days ahead, at noon)
• If it's full, join the waitlist? (default yes)
• Any dates to skip? (default it asks the day before)
• Cancellation window to remind you of? (default 12 hours)

Needs

• Login to your booking app or studio site. The agent books and cancels classes only; it doesn't buy credits or change payment details.
• Your list of usual classes. Fallback: tell it each week ("book Tuesday spin").
• If your assistant can only chat: it can remind you the minute registration opens and name the exact class to tap; you book it.

Build

1. Read: each day, list classes from your list that are exactly 7 days out.
2. Decide: skip any date you marked off. If the class costs more than usual, or the time or studio changed, ask before booking.
3. Deliver: at the moment registration opens, book. If it's full, join the waitlist.
4. Deliver: text you the class, time, place and cancellation deadline.
5. Memory: keep a list of bookings made so the same class is never booked twice.
6. If the site blocks the agent as a bot, text you the link to book by hand.

Check

• Dry run: list next week's matching classes and the exact minute each opens, without booking.
• A week with every class skipped sends nothing.
• Confirm the first real booking shows up in your app.

In my version

• Instinct, a texting AI agent, logged in to my ClassPass. Registration opens 7 days ahead at noon.

Ask first

• Which venues? (default the 3-5 biggest within an hour of you)
• Which artists should always be flagged? (default none; add as you go)
• Flag kid and family events? (default yes)
• What to skip? (default sports)
• How should alerts reach you? (default push notification; fallback email)
• What day and time? (default Monday, early morning)

Needs

• Each venue's public events page. If one won't load, use a ticketing site's page for that venue and say which you used.
• A watchlist of artists.
• One tracker file in your cloud drive. Connecting the drive lets it read and rewrite that file; it does not open other files.
• If your assistant can only chat: ask weekly "any new shows at [venues]?" and paste last week's list; it tells you what's new.

Build

1. Read: fetch each venue page; pull show name, dates and link. Read the tracker.
2. Decide: a show is "new" if its name and venue aren't in the tracker; a "date change" if they are but dates differ; "removed" if it's gone from the page while its dates are still in the future.
3. Decide: flag a new show if the artist is on the watchlist or it's a kid or family event (Disney On Ice, Sesame Street, a kids' circus). Drop sports.
4. Deliver: rewrite the tracker with every current show, its flags, and a "last checked" date.
5. Deliver: if any show is new, send one alert listing them, flagged first. Only date changes, or nothing new: no alert.
6. Deliver: for each flagged show, add a to-do "New show: [name] at [venue]" with dates and link. If to-dos can't be created, email yourself instead.
7. Memory: the tracker is the memory. The first run only builds it and sends no alert.
8. Cost: about 4-5 runs a month, a few page fetches each.

Check

• Run with alerts off. Every tracker row must link to a real venue page.
• Delete one show from the tracker and run again; it should come back marked new.
• A week with no changes ends with no alert.

In my version

• A Claude scheduled task, Mondays at 4am Pacific.
• Venues: Sphere (Las Vegas), Levi's Stadium, Shoreline Amphitheatre, SAP Center.
• About 20 pop artists on the watchlist. The tracker is an HTML file in my Google Drive; to-dos go to a to-do app I built.

Ask first

• When should the draft be ready? (default the evening, three days before game day)
• Where do the schedule and roster live? (default one shared spreadsheet, one row per game)
• Which details go in the email? (default time, opponent, home/away, jersey, snacks, field clean-up, absences)
• How do you open and sign off? (default "Hi All!" and your first name)
• Does the league email the schedule? (default yes; used as a cross-check)

Needs

• A schedule with one row per game date. Fallback: paste the season schedule into a doc once.
• A roster with parent emails. Fallback: a pasted list; report any family with no email.
• Your sent mail, to see whether you already sent it. Connecting email lets it read and create drafts; it never sends to parents.
• If your assistant can only chat: paste this week's row and ask for the email, then copy it into your mail app.

Build

1. Read: search your sent mail since last week's reminder day for this week's reminder.
2. Decide: if it's there, email yourself "Already sent, no draft made" and stop.
3. Read: find the next game in the coming 7 days; pull its fields. Read the roster emails fresh every time.
4. Read: check league emails from the past 10 days; note any time, field or opponent that differs from the sheet.
5. Decide: include the clean-up and absence lines only when filled. No game in the next 7 days: no draft; tell yourself.
6. Deliver: create the draft to every parent email. Short bullets. Never send it.
7. Deliver: email yourself the full text, any guesses, and any mismatch. You review and press send.
8. Cost: about 4 runs a month.

Check

• Run it today with drafting off. The output must match next game's row field by field.
• In a copy of the sheet, change one game time; the note to you must flag the mismatch.
• A bye week produces only the "no game" note.

In my version

• A Claude scheduled task with Gmail, Google Drive and Calendar connected, Wednesdays around 8pm.
• A Google Sheet with schedule and roster tabs for my daughter's under-6 team, about 10 families. Schedule emails come from the league's registration site.

Ask first

• Which past posts show your voice best? (default your 3 most recent)
• Section format? (default Problem / Solution / Why it works / Try it out)
• Which reader personas, in what priority? (default hiring manager, LinkedIn connection, former boss, close friend)
• LinkedIn format? (default 2-3 lines plus the link, never the full post)

Needs

• The build session, or your notes on what you built. Fallback: a 5-line summary from you.
• Three or more past posts. Fallback: any paragraph you've written.
• If your assistant can only chat: paste the notes and the posts; everything below still works.

Build

1. Read: the session or notes, the past posts, and any saved style notes.
2. Decide: pick the one problem the build solved. Open with a concrete scene, not a statement. Include one honest downside.
3. Deliver: the draft, in your sections.
4. Decide: each persona reads it and gives at most 3 notes (what to cut, what's unclear, what kept them reading), each quoting a line. Rank notes by persona priority.
5. Deliver: a revised draft applying the notes, plus the notes it skipped and why.
6. Deliver: the LinkedIn blurb. You publish both yourself.
7. Memory: after you publish, compare your final post to the draft and save what you changed as style notes.

Check

• Run it on the notes behind an old post. The draft should read closer to that post than to generic AI writing, and every persona note must quote a line.
• The revised draft is shorter than the first, not longer.

In my version

• Claude Code for the build, then three Claude skills: draft, four-persona critique, LinkedIn blurb. My newsletter is Lisa's Lab on Substack.

Ask first

• 5-10 emails you've sent that sound like you? (default your last 10 sent)
• Usual opening and sign-off? (default "Hi [Name]," and your first name)
• Words you never use? (default delve, navigate, "I hope this finds you well")
• One warm touch per email? (default yes: a ":)", a "!!" or a short personal aside)

Needs

• Sample emails. Fallback: describe your style in 3 lines.
• If your assistant can only chat: paste the thread and say who it's to. That is the whole routine.

Build

1. Read: the thread or the ask, and who it's going to (teacher, friend, professional contact).
2. Decide: put the ask in the first 1-2 sentences. Short paragraphs. Bullets only for 3 or more action items. Fix typos silently.
3. Deliver: the draft. You send it.
4. Memory: when you edit a draft before sending, save what changed as a style rule.

Check

• Draft a reply to an old thread and compare it to what you actually sent. The ask should land in the same spot.

In my version

• Claude with Gmail connected, plus two email-voice skills I keep refining from my own edits.

Ask first

• Simplified or Traditional characters? (default Simplified)
• Tone? (default formal but warm, like a school office letter)
• Who checks it before it goes out? (default a parent volunteer who is a native speaker)
• Keep names and links in English? (default yes; dates in Chinese format)

Needs

• The English email.
• 2-3 past translations a native speaker approved. Fallback: none at first; save the reviewer's changes as you go.
• If your assistant can only chat: this already is a chat routine; paste the email and ask.

Build

1. Read: the English text and the approved examples.
2. Decide: match the examples' greeting, sign-off and terms, using a short glossary of event and group names. Translate meaning, not word for word. Keep every date, time and amount exact.
3. Decide: flag any idiom or joke that doesn't carry over, with a suggested swap.
4. Deliver: the Chinese version, then the flagged spots. The reviewer approves before it is sent.
5. Memory: add the reviewer's changes to the glossary.

Check

• Translate last month's email and compare it to the approved version. Every date, time and amount must match exactly.

In my version

• A Claude skill, for the parent organization at my kids' school (I'm president). The Chinese goes out in the same email as the English.