Scope a feasibility spike live for a problem I describe.

Here's the problem to scope live: Trestlebridge Estimating sells bidding software to general contractors. The interviewer describes a new feature idea: an AI that reads an uploaded blueprint PDF and drafts a materials takeoff, lumber, concrete, rebar counts, for a contractor to review before bidding. Dax Whitfield is the senior estimator who'd actually use it. Rutger Vandermolen is the candidate walking through how he'd scope the spike, live, in the room.

How to answer this, stage by stage

Nobody is scoring whether the invented product is realistic. They're scoring whether you can turn a cold problem into a concrete anchor and a real risk inside about ninety seconds.

Let's learn

Picture this product before anyone has touched a single setting on it: a blueprint PDF goes in, a materials count comes out, and somebody's whole afternoon depends on whether that count is right.

Today at Trestlebridge, an estimator reads a blueprint page by page and tallies lumber, concrete, and rebar by hand, about six hours for a typical residential bid, across roughly eighteen bids a month. With Takeoff Draft, a full count comes back in about twenty five minutes, ready for an estimator to check against the plan.

The five letters, held up as one page. Anchor is the step this exercise is really testing.

Here's the turn: the first version of Takeoff Draft, sketched fast to get something in front of Dax, produced one single, all-or-nothing count for the whole blueprint set. If any one page got misread, an estimator's only option was to throw out the whole draft and redo the takeoff from scratch, the same six hours it was supposed to save.

At its worst, a hand-marked revision cloud on one page gets read as a real wall, the count comes back over-ordered on lumber and concrete, and the whole draft has to be scrapped and redone by hand anyway, wiping out the entire time saved.

What I would leave alone: I wouldn't build automatic reconciliation across conflicting drawing versions on day one, since that's a genuinely harder problem and proving the core idea doesn't depend on solving it first.

The lesson: a spike for an AI feature isn't really testing whether the model is usually right. It's testing what happens to the person's afternoon on the day it isn't.

Now here is the same thing as a story

The short version above is what you'd say scoping this live, in the room. Read this one for what it actually looked like the week two bad bids in a row changed how Dax used the tool.

Dax Whitfield has drafted materials takeoffs for eleven years, and can usually tell from a blueprint's title block alone whether a job is going to be a headache.

The first step is where a hand-marked revision cloud either gets read correctly or doesn't, and everything downstream depends on it.

When Takeoff Draft launched, Dax reviewed its first counts carefully against the blueprint, page by page. They matched, every time, for the first forty bids. By bid forty-one, he was skimming the summary total and moving straight to submitting the bid.

Four things the anchor decision actually needs, made inspectable instead of just described.

Two misreads, two days apart, and a habit changed that nobody had to announce out loud.

Bid forty-two had a revision cloud circling a doorway change, and Takeoff Draft counted it as an extra section of wall. The over-order wasn't caught until material arrived on site, eighteen thousand dollars of lumber nobody needed. Two days later, bid forty-three had the same problem on a different page, twenty two thousand dollars this time.

Same mistake, two very different afternoons, depending entirely on which of these two designs was in front of Dax.

Dax didn't complain or file a ticket. He started tracing every blueprint page into a cleaner drawing program before ever uploading it to Takeoff Draft, manually removing every revision cloud himself first, an extra hour added back onto every single bid.

The real question was never whether Takeoff Draft's model was usually accurate. It was whether the product survived the specific, predictable day it misread a revision cloud.

One branch was worth solving immediately. The other was worth naming and setting aside on purpose.

When Takeoff Draft was first designed, someone said, "let's just ship one clean count per blueprint, simplest thing that could work," and it sounded reasonable, since the clean pages were never the problem.

Rerun the same two bids with the page-level anchor in place: bid forty-two's revision cloud gets a low-confidence flag on that one page, Dax fixes just that page in about twenty five minutes, and the over-order never reaches a material order at all. Bid forty-three ships the same way. Two five-figure mistakes become two twenty five minute corrections.

What I'd tell myself, watching Dax quietly start tracing blueprints by hand before ever touching the tool: the six hours it saved were never really saved. They were just borrowed, waiting for the first page it misread.

SPARK, the anchor that had to survive its own bad dayNot a script for over-engineering a first version. SPARK is what tells you exactly which one decision a live scoping answer actually needs.

The recap, one line per letter: situation is Dax's six-hour manual tally, payoff is ending that tally without creating blind trust in one total, anchor is a page-level draft with per-page confidence and flags, risk is a misread revision cloud costing a real over-order, and keep out is skipping cross-version reconciliation on day one.

And if you want to be sure it really works, try it somewhere elseSame five letters, a translation service instead of a construction bid. Different flip family entirely, the same missing anchor.

Idris Vantol runs product at Anthem & Oak Translation Services, where a newer feature localizes marketing copy into a target language for an in-country reviewer to approve. Mapped onto SPARK: situation is a reviewer today reading every line of a manual translation before it ships. Payoff is ending full line-by-line reading for copy that's already reliably good. Anchor is showing which specific phrases were rewritten for cultural fit versus translated directly, not just a single fluent block of text. Risk is copy that reads perfectly natively while quietly changing the actual meaning of one claim, a review habit fluent-sounding text makes people drop fastest. Keep out is fully automating tone calibration across every regional dialect, a much bigger problem than proving the core idea needs solved. But the flip here is an over-trust one, not an input flip: once localized copy started reading fluently in-language, in-country reviewers who used to check every line stopped checking nearly any of it, since fluent had quietly become their whole definition of correct.

A different flip entirely: not a person performing more for the machine, but a person checking less because the machine performs so well.

Swap the trigger and it still runs.
Speed: an interviewer caps you at sixty seconds. Say "one person, one anchor decision, one failure mode, in that order," and stop there.
Cost: no time to build a real page-level prototype before the room moves on. Say so honestly, and sketch the anchor on paper as the thing you'd build first with any spare hour.
The model turns out to be excellent on messy real blueprints, for real: if revision clouds stop being a problem entirely, that's still worth verifying with real data before removing the per-page flag, since one clean quarter isn't the same as a guarantee.

Where people run it wrong.
They open a live scoping question by asking for more requirements instead of committing to one concrete case.
They list every possible test instead of naming the one anchor decision the spike actually turns on.
They design the happy path first and treat the failure mode as an afterthought instead of the actual design target.

How to use it live. The moment an interviewer hands you a problem cold, ask yourself: who's the one person doing this by hand today, and what's the one decision that has to survive their worst day with the tool? Say both out loud in the first breath, and the rest of the scoping conversation follows on its own.

Flashcards (tap any card to flip it)

Check yourself Score: 0 / 0