What question should a feasibility spike answer, and what should it explicitly not try to answer?

Solmark Agritech is building CropSentry, which reads a phone photo of a tomato leaf and flags likely disease. Solveig Amundsen is the AI PM deciding what a two-week feasibility spike is actually for, after a six-week attempt to prove it works on everything at once produced no answer at all.

How to answer this, stage by stage

Nobody is scoring whether you can list every possible thing a spike could test. They're scoring whether you can say, in one sentence, which single thing it actually has to.

Let's learn

Every day, Tomasz Wysocki walks about three acres of tomato fields on foot, spending roughly 25 minutes an acre checking leaves that look off, mostly telling early blight apart from a common, harmless leaf spot that looks almost identical in its first days.

Say we build a tool that reads a phone photo of a leaf and flags likely disease in seconds, instead of a scout walking every row by eye.

Step four only happened because Tomasz still walks fields out of old habit. Nothing in the system made it happen.

Here's the turn: the extra false flags on hard cases were never the real problem. The real problem was that Solmark's first spike tried to prove CropSentry worked across all 12 crops it might eventually support, and six weeks in, still had no answer, while the one confusion that could actually cost a field, blight versus a harmless look-alike, had never been tested on its own.

At its worst, CropSentry ships trained mostly on well-lit sample photos, quietly misreads a real blight case as harmless under ordinary midday glare, and by the time symptoms are obvious enough for a human to catch by eye, a whole section of the field is already lost.

What I would leave alone: whether CropSentry eventually supports all 12 crops is a real question, just not this spike's question. Answering it here would have meant never finishing at all.

The lesson: a spike that tries to prove everything proves nothing on any deadline. A spike scoped to one costly confusion proves something in two weeks.

Now here is the same thing as a story

The short version above is what you'd say defending the narrower scope in a review. Read this one for what almost missing the near miss actually felt like.

Solveig Amundsen had run three model launches before CropSentry, each time waiting for a clean, unambiguous "it works" before greenlighting a build.

The first spike attempt was ambitious on purpose: test all 12 crops Solmark eventually wanted to support, every growth stage, every lighting condition, before saying anything definitive. Six weeks in, the team had covered about 40 percent of that plan, with a growing spreadsheet of partial results and no clear answer for leadership.

Only one confusion sat in the corner that actually mattered. The six-week plan was testing all four equally.

Tomasz, an agronomist who still walks fields out of old habit even with CropSentry running, pulled up a batch of 50 leaves the tool had flagged "low risk" and checked them by eye anyway, the way he'd done long before any app existed.

One leaf stood out: CropSentry had confidently called an early blight case "harmless leaf spot, low priority." Tomasz caught it purely because that particular row happened to be on his usual walking route that week, not because any part of the system flagged it for review.

Drawn to scale on purpose. One of these mistakes is loud and cheap. The other is silent and costs a field.

Solveig scrapped the 12-crop plan and scoped a new, two-week spike around exactly one question: could CropSentry reliably tell blight from leaf spot, specifically under the lighting conditions field photos actually get taken in.

Four things, decided before day one. Everything the six-week plan lacked, in one page.

The real question was never whether CropSentry could tell diseases apart in general. It was whether it could tell apart the one pair that actually costs a farmer a field, in the actual light a phone camera sees at two in the afternoon.

Fourteen days, one confusion, four lighting conditions. Nothing else was allowed onto the calendar.

When the first plan was proposed, someone said, "let's be thorough and test the whole crop lineup before we commit to anything," and it sounded responsible, since nobody wanted to greenlight something narrow and get blindsided later.

Saying what a spike won't answer turned out to matter as much as saying what it will.

Rerun the six weeks with the narrow scope already chosen: the 54 percent midday-sun gap surfaces on day ten, against a tightly built test set, instead of surfacing whenever the next scout happens to walk that particular row by habit.

What I'd tell myself, hearing how close that field came to being lost by luck alone: thoroughness that never finishes isn't caution. It's just a different way of never actually finding out.

PICK, the tradeoff that scopes the questionNot a script for testing less. PICK is what tells you exactly which one thing has to be tested first.

The recap, one line per letter: position is the one confusable pair under real light, impact is a wasted spray versus a lost field, cost asymmetry is $140 against $8,400 an acre, and kill criteria is 85 percent in every lighting condition, which midday sun failed at 54.

And if you want to be sure it really works, try it somewhere elseSame four letters, a pharmacy counter instead of a tomato field. The costly confusion is still the one thing the spike has to nail.

Delphine Okwuosa runs product at Larkspur Pharmacy Group, testing ScriptScan, a tool meant to flag a photographed prescription as likely altered before a pharmacist fills it. Mapped onto PICK: position is whether ScriptScan can tell a genuinely altered prescription apart from a legitimate one written in unusual handwriting, not whether it can catch every conceivable fraud scheme that might exist. Impact is a delayed, embarrassed patient with a legitimate script versus a controlled substance fraud that goes through unflagged. Cost asymmetry is a delay costing a patient an inconvenient extra call to their doctor, against a missed controlled-substance fraud costing a real health and legal risk that can't be undone once the medication leaves the counter. Kill criteria is catching at least 95 percent of confirmed past fraud cases without flagging more than a small, stated share of legitimate ones.

A different counter, the same shaped tradeoff: the costly error is the one that can't be quietly corrected after the fact.

Swap the trigger and it still runs.
Speed: an interviewer caps you at sixty seconds. Say "name the one costly confusion, test it under real conditions, set the bar around the expensive error," and stop.
Cost: no time to build four lighting conditions into the test. Say so honestly, and test the single worst condition you already suspect, rather than an average across easier ones.
The model really does handle it, for real: if every lighting condition clears the 85 percent bar, that's a genuine green light, and saying so plainly is what makes the method trustworthy instead of reflexive doubt.

Where people run it wrong.
They scope a spike to "prove it works," which never has a stated finish line.
They test on the easiest, cleanest version of the data and report an average that hides the case that matters.
They treat every wrong answer as equally bad instead of naming which one actually costs something.

How to use it live. The moment you're asked what a spike should answer, ask yourself: which two things does this model most need to tell apart, and which mistake between them actually costs money? Scope the spike to exactly that, and say out loud what it deliberately leaves for later.

Flashcards (tap any card to flip it)

Check yourself Score: 0 / 0