Design a two-week spike to test whether an AI feature is viable.

Casterbridge Realty Group is building RenoScope, a tool that reads a listing's photos and estimates renovation cost, so property investors can triage which listings are worth a bid before ever flying out to see one. Bartholomew Kase is the AI PM who has two weeks to prove it works before anyone commits real budget to it.

How to answer this, stage by stage

Nobody is scoring whether you can fill fourteen days with tasks. They're scoring whether the order you chose protects against the one mistake you can't take back.

Let's learn

Say we build a tool that reads a real estate listing's photos and estimates how much renovation the property will need, so an investor can decide which listings are worth pursuing before ever visiting one.

Before RenoScope, an acquisitions analyst spent about 20 minutes per listing eyeballing renovation scope from photos, across roughly 50 listings a week, nearly 17 hours weekly just triaging which properties were even worth a closer look.

Four things, pulled from real closed deals, before any accuracy number means anything at all.

Here's the turn: the team's first attempt at proving RenoScope out wasn't badly built, it was badly ordered. Everyone was ready to spend the whole two weeks wiring a rough model into the live listing feed and building a clean dashboard, before anyone had checked whether its cost estimate was even in the right neighborhood on a real, messy set of past deals.

At its worst, RenoScope ships on the strength of a demo built only from newer homes, an investor bids on an older property based on a rosy estimate, and a hidden foundation or wiring issue turns a projected 20,000 dollar flip into a six-figure loss, the kind of miss that ends the whole program's credibility in one deal.

What I would leave alone: for newer, post-1980 properties, I wouldn't hold the launch hostage to a perfect number on the oldest homes. RenoScope can ship for the segment it's already proven on while the older-home gap gets fixed separately.

The lesson: a two-week spike doesn't fail because the tasks were wrong. It fails when the order lets an unrecoverable mistake happen before a recoverable one gets the chance to catch it.

Now here is the same thing as a story

The short version above is what you'd say proposing the redo to leadership. Read this one for what the audit that forced the redo actually felt like.

Bartholomew Kase had been at Casterbridge two years, long enough to know which investors trusted a number and which ones wanted to see the receipts.

The first RenoScope attempt moved fast: a rough model, ten sample estimates, and a dashboard mockup ready to show at the next investor call. Nobody had checked the ten estimates against anything real yet, but the demo looked sharp, and the room was excited.

Two ways to be wrong. Only one of them can be quietly fixed afterward.

Persephone Ilić, who runs operations audits at Casterbridge, pulled ten of the sample estimates at random and checked each one against the actual signed contractor invoice from the same property, a habit she'd built long before RenoScope existed.

One property stood out: RenoScope estimated 18,000 dollars in cosmetic repairs. The actual signed contractor invoice was 142,000 dollars, almost entirely from a foundation issue no listing photo had ever captured. Persephone brought the number to the next planning meeting, not as a complaint, just as a fact nobody had checked yet.

The dashboard build sat exactly where it shouldn't have been first: hard to undo, and unblocking nothing.

Bartholomew scrapped the plan to spend the two weeks polishing the dashboard and instead ranked the actual sequence: assemble a real golden set first, since nothing else could be trusted without it.

The live build sits last on purpose. Everything before it exists to make sure it's worth doing at all.

The real question was never whether RenoScope could produce a number. It was whether anyone had ordered the two weeks so the costliest kind of wrong got caught before an investor ever saw it, not after.

Five items, in the order that matters. The last one is deliberately last.

When the first plan was drawn up, someone said, "let's get the dashboard looking right, that's what'll sell it in the room," and it sounded reasonable, since a sharp demo really had gotten leadership's attention before.

Fourteen days, four real checkpoints. No day left open for "and also polish the demo."

Rerun the same two weeks with the ranked plan already in place: the 142,000 dollar gap surfaces on day nine, against a private golden set, at a cost of about 14,000 dollars, instead of surfacing after a real investor had already bid low on a hidden six-figure problem.

What I'd tell myself, hearing Persephone's number land in that meeting: a sharp demo proves the model can produce a confident answer. It was never proof that the confident answer was the right one.

ORDER, the sequence that protects the mistake you can't undoNot a script for slowing every spike down. ORDER is what tells you which test has to come first, and which one can safely come last.

The recap, one line per letter: outcome is safe listing triage, reversibility is a live bad number versus a private one, dependency is the golden set unlocking everything else, evidence is a same-day cheap pass before full commitment, and rank is data, accuracy, worst case, trial, then build, in that order.

And if you want to be sure it really works, try it somewhere elseSame five letters, a fishing dock instead of a listing photo. The costliest mistake still isn't the one everyone expects.

Sigrun Haugland runs product at Nordveil Fisheries Co-op, testing CatchWeigh, a tool meant to estimate a haul's total weight from a deck photo, to speed up dockside sorting before the truck scale confirms it. Mapped onto ORDER: outcome is pricing a haul fairly and fast without slowing offloading. Reversibility is a wrong private test number, easy to rerun, versus a wrong price quoted to a fisherman at the dock, which is hard to undo once trust in the tool is gone. Dependency is needing real scale-weighed hauls matched to deck photos before any accuracy test means anything. Evidence is a same-day check against 20 archived photos and their known scale weights, before committing the two weeks. Rank is assemble the weighed golden set, test single-species accuracy, then deliberately test mixed hauls where fish overlap in the photo, then one narrow trial with a single dock crew, with the register system left untouched until all of that holds up.

A different dock, the same shaped decision: single-species accuracy alone was never going to be the whole answer.

Swap the trigger and it still runs.
Speed: an interviewer caps you at sixty seconds. Say "rank by which mistake can't be undone, put the live build last," and stop.
Cost: only a few days available, not two weeks. Say so honestly, and compress to the golden set and the accuracy test alone, skipping the trial rather than skipping the ordering discipline.
The model really is ready, for real: if the worst-case test also holds up, that's a genuine green light to build the integration next, and saying so plainly is what makes the ranked plan trustworthy instead of reflexive caution.

Where people run it wrong.
They order a spike by what's fastest or most exciting to build, not by which mistake is hardest to take back.
They test only the easy, representative case and report an average that hides the segment likely to fail.
They let the dashboard or integration sneak into the two weeks before the number behind it has ever been checked.

How to use it live. The moment you're asked to design a two-week spike, ask yourself: which possible mistake, if it reached a real user first, could I never fully undo? Put whatever prevents that mistake first, and the rest of the order tends to follow.

Flashcards (tap any card to flip it)

Check yourself Score: 0 / 0