How do you teach users what the AI can and cannot do without a manual?

Ridgeline Stamping runs a press line that turns sheet steel into car door panels. Priya Deshmukh has run that line for nine years. GlintCheck is the camera mounted over the last press, checking every panel as it comes off.

How to answer this, stage by stage

A student who reads only this section, and reads it out loud once, should be able to answer the question cold.

Let's learn

For nine years, Priya Deshmukh walked the press line at Ridgeline Stamping at six in the morning, and could spot a hairline crack in a stamped door panel before the machine even finished its cycle.

GlintCheck is a camera mounted over the last press. It checks every panel the second it's stamped, for cracks, misaligned holes, and warping, the three defect types it was trained on.

Before GlintCheck, Priya and her crew spot-checked one panel in every ten, about 400 of the 4,000 stamped in a shift. They caught most of what they were looking for. Now GlintCheck checks all 4,000, and catches nearly every crack and misaligned hole, over 97 out of 100.

At its worst: three weeks after a coating supplier switch, a batch of 1,200 panels with coating streaks shipped past GlintCheck, which had never been shown that pattern, and past every operator, who had learned to trust a clean scan as the whole story. A routine skip-level quality audit caught it, not the line.

What I would leave alone: crack and hole detection do not need any of this scaffolding. GlintCheck matches an expert's judgment on those almost every time, tested across more than 50,000 panels. Adding caution there would only slow the shift down for no real gain.

The lesson: a tool that checks everything is not the same as a tool that checks everything you care about. If the onboarding never says which is which, the operator has no way to know, and neither did we, until an audit found out for us.

Four items, and the second one is the one the old ten-minute walkthrough never had.

Now here is the same thing as a story

The short version above is what you'd say defending the design in an interview. Read this one for how it actually happened on the floor.

The press line at Ridgeline Stamping starts loud at 5:45, when the first coil feeds through and the stamps begin their rhythm. Priya has run that rhythm for nine years. Ask her to find a hairline crack in a stack of a hundred panels and she'll have it before you've finished asking.

Before GlintCheck, one in ten panels got a real human look. The other nine rode on Priya's memory of what usually goes wrong.

GlintCheck arrived in February. For the first few months it was a genuine relief: every single panel got checked, not one in ten, and Priya's crew could spend their attention on the panels the camera actually flagged instead of guessing which ones to sample.

The first-shift walkthrough new operators got was short: log in, watch a demo scan run clean, see a green checkmark, start the line. Nobody asked what the checkmark didn't cover. Nobody had been given a reason to ask.

Same camera, same panel. The difference is only in what the operator was ever told to watch for beyond the screen.

Then Ridgeline switched coating suppliers, for a cheaper, faster-drying finish. Nobody flagged it as a training-data event, because nobody thought of GlintCheck as having training data at all. It was just "the checker."

Coating streak sat in the one corner nobody had drawn on the map yet: subtle, and almost never caught.

For three weeks, the streaks kept coming, and GlintCheck kept passing every panel clean, because a streak was never in its trained set. Then a skip-level quality audit, the kind that happens twice a year, pulled a random sample and found 1,200 panels already shipped downstream with visible coating streaks.

We did not lose a camera's worth of accuracy. We lost three weeks nobody was watching the one thing the camera was never built to watch.

The fix wasn't a smarter camera. It was a four-minute redesign of that first shift.

Step two is the one that used to be missing entirely. A photo of a real miss teaches the edge faster than any sentence about it.

Four minutes, four milestones. Nothing here needed a manual, only a screen and two real photos.

Now every new operator sees one real photo of a caught crack, and one real photo of a coating streak GlintCheck waved through, side by side, on their very first login. They're told, plainly: this camera has a list of things it looks for, and this isn't one of them yet. If something looks wrong and the screen says clean, log it anyway.

I built the old walkthrough because it was fast to run and easy to schedule everyone through on their first day. It took watching 1,200 panels ride past two separate sets of eyes, human and camera, before I understood that "checked" and "checked for what you care about" are not the same claim, and our onboarding had only ever taught the first one.

SPARK, the design behind the four minutesNot a checklist for operators. SPARK is what forces the design itself to survive the day it's wrong.

The recap, one line per letter: situation is the trusted checkmark, payoff is noticing what got skipped, anchor is the catch-and-miss first screen, risk is a genuinely new defect type, and keep out is the confidence overlay we're saving for later.

And if you want to be sure it really works, try it somewhere elseSame five letters, a denim mill instead of a stamping plant. A completely different fabric, and this time the edge is a dye lot, not a coating.

Aldergate Mills runs LoomSight, a camera system over its looms that checks woven denim for flaws, snags, misthreads, tension lines, as it comes off each roll. Hollis Vance is the quality technician who signs off on every roll before it ships to a cutting house.

Mapped onto SPARK: situation is Hollis walking a physical roll by hand under a light table before LoomSight existed, catching snags by touch as much as sight. Payoff is the habit of noticing a weave pattern LoomSight has never priced in, instead of trusting a clean scan as final. Anchor is the same catch-and-miss first session, but built from fabric photos: one real snag caught, one real irregularity from a brand-new dye lot that LoomSight waves through because it has never seen that pattern before. Risk is a new dye supplier shipping a pattern the model has never priced in, exactly the coating-switch problem in a different material. Keep out is a full thread-count overlay, which is real information Hollis doesn't need on day one.

The branch that matters most is the second one: a pattern nobody trained the model on gets routed to a person, not waved through silently.

Swap the trigger and it still runs.
Speed: an interviewer caps you at sixty seconds. Say "show a catch, show a miss, give a log-it rule," and stop there.
Cost: there's no budget for a proper first-shift redesign this quarter. Say so honestly, and start with a single laminated card taped by the tablet showing one miss photo, since even a low-effort version beats none.
The model gets better, for real: if GlintCheck's overall accuracy climbs to 99 percent, that's still not a reason to drop the miss-example, a rarer failure is a more dangerous one to have nobody watching for.

Where people run it wrong.
They write a manual nobody opens, instead of building the lesson into the product's own first moment.
They quote one overall accuracy number and assume it teaches the boundary, when it hides exactly which things were never in scope.
They treat "the model got better" as a reason to remove the caution the first design earned, instead of a fresh reason to re-check it.

How to use it live. When someone asks how you'd teach capability without a manual, ask yourself one question first: what's the one photo of a miss I'd want a brand-new operator to see before their first real panel. Say that photo out loud, not a policy about training materials.

Flashcards (tap any card to flip it)

Check yourself Score: 0 / 0