01Problem

Smooth onboarding vs context gathering

An agent can't book well without context, and every signup question gives me a reason to quit.

  • Muse asked nothing and read my $1,500 as flights only, so its picks came in 67% over.
  • People skip a long setup interview, and questions about income feel invasive.

How much the agent asks before it books

Connect and goMuse, CA$2,500 on a CA$1,500 budget
Hindsight2 questions at bedtime, right before it spends
Interview at setuppeople skip it or quit
02Demo

Hindsight

Hindsight asks what the agent can't guess right before it spends, then checks the work against real records.

Open the demo
03Principles

Preferences are prices

An agent can't pick for me until it knows what my time is worth in dollars.

  • A founder and a saver can both say they prefer direct flights, but only a dollar amount shows who will pay $180 for one.
  • My bedtime answers set the price, and each Change I make moves it.

What skipping a 2 hour connection is worth

Founder
+$107, Muse's direct +$180, the brief's direct
Saver
Me
$0$50$100$150$200$250
pays itunsure, so it askstakes the directkeeps the connection

Ask at the moment of commitment

A few questions right before a big purchase feel like care, the way my bank's 2FA prompt does before a large transfer.

  • Setup stays one tap, and the questions come at bedtime.
  • If I fall asleep without answering, it uses the default it showed me.
Confirm a CA$1,900 payment
ApproveDeny
Hindsight 11:49 PM
book my reading week trip, toronto → nyc → boston and back, oct 10 to 18, under $1,500
reading it back: toronto → nyc → boston → toronto, oct 10 to 18, window seats. is the 1,500 cad and does it cover hotels too?
cad, everything
ok. flights + hotels + train come to about 1,900 at the cheapest. that's 27% over, inside the 30% you said. i'll go ahead unless you say no
one thing i can't price for you. there's a 6:53 am train, or a 9 am one with a comfier split for +150. if you don't answer i'll take the 6:53

I didn't reply, so it became a card in the morning.

The agent decides anything that can be undone

It books those right away and lists them in the morning digest.

  • The US 24-hour rule lets me cancel any flight on this trip for free within a day of booking.
  • Anything costly waits for me unless a rule I signed covers it.
  • When the cheap option can't be undone, Hindsight books a refundable one if it fits my 30%. The Boston dorm had none that fit, so nothing got booked.
Knows my preferenceUnsure Free to undo
Does it, tells me in the digest Answered Instinct's 5 questions Turned down the 14 hour bus Booked both flights, window seats Booked 3 nights in NYC
Books its best guess and flags it 6:53 train, or 9:00 for +CA$150?
Costly to undo
Does it only inside a rule I signed a $60 bus seat, under my $150 cap
Waits for me Book the Boston dorm? Can't cancel

Checked against real records

Every claim gets checked against a record the agent didn't write.

  • The airline's booking page and the card charge show what was actually bought.
  • A second agent reruns the search in the background, and I only hear from it when it disagrees.
Instinct said, 5:07 PM
"transport total: ~CA$465-480"
Fare quote, 6:27 PM
CA$673 for the same flights
CA$193 low, caught before booking
Muse said, 5:02 PM
"Found 3 itineraries under your budget"
My answer, 5:14 PM
"Cad and budget for everything"
flights only, about CA$2,500 with hotels
Booked, 1:15 AM
Air Canada 9 PM Sat Oct 10, YYZ → LGA, window seat
Airline booking page
same flight, seat, fare and total in CAD
matches

No new app to manage

Hindsight starts as one message in the chat I already use.

  • The audit opens as a page from a link card in that thread.
HindsightToday 7:02 AM
morning! booked your trip, about $1,900 all in
2 things need you before 9. the train time and the boston dorm, which i won't book without you
See what the agent did11 steps overnight, 2 need you
04Design

Five layers of detail

Each layer down shows more detail, and most people stop at the notification.

Ask about conflicts before bed

Hindsight asks which limit bends first before I sleep, so a 3 AM conflict already has an answer.

  • It sends one message with a deadline when nothing fits, then books once the deadline passes.
  • Keep is selected unless I tap Change.
11 PM1 AM3 AM5 AM7 AM
11:49 PM"budget can go over, about 30%"
3:00 AMunder budget means a 6 hour layover
7:00 AMone card
Under CA$1,5006 hour layover
Inside CA$1,950, bookedshorter routing, over budget
KeepChange
05Next

What comes next

When I send a command and go to sleep, there's no time to ask me anything, so the next version needs to know more before I send it.

Notes

How it works

Hindsight sits between me and my agents. My preferences live in Hindsight instead of inside the agents, so the check stays independent.

  • It stores what my time and risk are worth.
  • It keeps rules for what it can undo, which limit bends first, and when to wake me.
  • It asks only when my answer would change the pick by a lot.
  • A checker with a fresh context compares each claim to a record.
Constraints I designed around
I'm asleep for 8 hoursRules I set at bedtime make the calls, and anything costly waits for me.
A fare expires in 20 minutesThe 24-hour rule makes the booking free to undo, so Hindsight books it and flags it.
The agent makes mistakesHindsight checks each claim against a record, and I can change any pick with one tap.
Instinct has no APIChange sends the redirect through the chat.
iMessage renders very littleThe audit opens from a link card.
Refundability is unknownHindsight treats it as non-refundable, and the card says so.
The cheap bed might sell outHindsight books a refundable bed if one fits my 30%. If none fits, it books nothing and the card says what waiting could cost. A hostel rarely holds a bed, so Hindsight never claims one is held.
Assumptions

The trip is booked 7 or more days out, directly with the airline, so the 24-hour rule applies. Hindsight can read booking emails and card charges. Prices are in CAD for one traveller.

Alternatives I rejected
A simulated version of meDigital twins predicted new money choices poorly (r = 0.20, Park et al. 2024).
A long interview at setupPeople skip it, and personal questions feel invasive.
Agents debating each otherTwo agents can agree on a wrong fare, so I check against records instead.
What's real and what's simulated

The agents' options, prices and quotes come from my Instinct and Muse chats on Sep 27, 2026. Hindsight's messages, bookings and checks are simulated from those options, and nothing was bought. The founder and saver rows, the 3 AM routing and the examples under What comes next are illustrative.

Sources
  • Jason Yuan's Mercury OS, for fading steps you aren't reading.
  • Dot at New Computer, for showing memory as short stories.
  • Hivemind's own site, for the type scale and hairlines.
  • NASA flight rules, for making decisions ahead of time instead of mid-emergency.
  • FAA alert colors (14 CFR 25.1322), for orange and red.
  • SBAR handoffs, for the single 3 AM message. A structured handoff program cut medical errors by 23% (Starmer et al. 2014).
  • Flighty, for times and terminals.
  • Malle et al. 2015. People blame a robot for not acting almost as much as for acting.
  • Dietvorst et al. 2018. Letting people adjust an algorithm raised its use from 32% to 73%.
  • Park et al. 2024. Simulated people predicted money choices poorly.
  • Shalowitz et al. 2006. Proxies guessed a patient's choice right 68% of the time.
  • Liu et al. People kept 78.7% of the settings from a few questions asked up front.
  • US DOT 14 CFR 259.5(b)(4), the 24-hour rule.