You can know the material cold and still lose the room. The usual ways are jumping straight to a solution, going deep on the wrong component, or never checking in. None of them is a knowledge problem. They are pacing problems.
Designing a system on your own and designing one in front of someone are two different skills. The first one is mostly knowledge. The second one is knowledge on a clock, with another person reading your face.
This is true whoever is watching: a design review, a whiteboard session with a new team, an incident bridge at three in the morning, a hiring panel. The audience changes. The failure modes do not.
When someone watches you design, they are judging how you handle ambiguity, time and disagreement at least as much as the diagram. The experienced get caught on pacing, not on knowledge. A handful of small habits fix it.
A Frame Everyone Borrows
The most widely used frame comes from Alex Xu's book System Design Interview: An Insider's Guide, published by ByteByteGo, and I am describing it in my own words. It has four steps. First, understand the problem and set the scope. Second, propose a high-level design and get buy-in. Third, a deep dive into the parts that matter. Fourth, a wrap-up.
Its time guidance is rough, and the page says so itself: three to ten minutes on the problem, ten to fifteen on the high-level shape, ten to twenty-five on the deep dive, three to five on the wrap-up. It also lists what the watcher is assessing besides the design: how well you collaborate, how you handle pressure and how you resolve ambiguity. Asking good questions is treated as a skill in its own right. The red flags it names are over-engineering, which it describes as pursuing design purity while ignoring trade-offs, narrow-mindedness and stubbornness, and jumping to an answer before the requirements are clear.
The frame was written for interviews. It works because every watched design has the same four moments.
The Risk Is Pacing
Here is the part I would underline for anyone with years behind them. The likelier way for an experienced engineer to fail this frame is skipping to step two or three, not lacking the knowledge. Fluency tempts you to draw the solution you already know. It feels like competence and reads, from the other chair, like not listening.
Because the clock is the real constraint, I find it more useful to think in ratios than in minutes. Roughly fifteen percent of the time to understand the problem, thirty to draw the shape, forty-five to go deep, ten to wrap up. That is my reading of the ranges, not a number from the source. If you get twenty minutes it is three, six, nine and two. If you get forty-five it is about seven, thirteen, twenty and five.
Six Habits
Ask before you draw, and ask the question that decides the shape. Most problems hide one fact that changes the whole answer. Is the load mostly reads or mostly writes? Is a stale answer acceptable for a minute, or never? Does a human check the output, or another system? Find that question early. It is also the quickest way to show you understood the problem.
Say the time budget out loud. "Can I take a few minutes to ask questions before I start drawing?" At the end: "We have about ten minutes left. Do you want me to go deep on one more piece, or wrap up?" Naming your own time management is a signal most people never send, and it gives the other person a way to steer.
Check in after the first shape. Once the boxes are up, ask: "Does that match what you had in mind, or is there a piece I am missing?" It is not a filler question. The answer tells you where to spend the deep dive.
Offer the deep dive as a menu. "I could go deep on how data is partitioned, how failures are handled, or how it scales. Where is it most useful?" You hand the choice to the person who knows what they are worried about, and you stop guessing.
Cut yourself off. If you notice you are three levels into a side detail, say so: "I could go further on the replication mechanics, but let me come back up. Is that the direction you wanted?" Catching your own drift out loud is a stronger signal than never drifting.
Never end on "that is my answer". Close with a boundary: "I would not call this finished. The piece I would pressure-test next is the retry path, and the failure I would watch for is a message sent twice." The source's own advice is not to call a design perfect. Naming a real failure mode before anyone asks is the same instinct as the pre-mortem in Design the Failure First, done live.
A Small Example, Invented
To see the habits at work, take a prompt I made up for this essay: design a service that sends reminders for unpaid invoices.
Ask. Is a reminder sent to one person or several? Does it stop when the invoice is paid, and how quickly must it know? What happens if the same reminder goes out twice?
A number, a conclusion, a decision. Suppose, for illustration only, there are two hundred thousand overdue invoices at any moment, and each gets one reminder a week. That is about a third of a reminder per second. This is not a throughput problem. It is a correctness problem: do not send twice, do not send after payment, respect time zones. The number tells you where not to spend your time, which is the point of doing it. I wrote about getting that kind of estimate on the page without false precision in Precision Is Not the Point.
Offer the menu. The design most likely to go wrong is the idempotent send, the stop-on-payment path, or the schedule. Ask which the person wants.
Close with the boundary. The thing to pressure-test is the gap between "paid" and "stop reminding": a payment that lands during the send.
What the Watcher Sees
A design review is also a test of whether the room can disagree. In Ego-Free, Blame-Free I wrote that a design review that ends four minutes early should make you suspicious. Everything above is a way of making disagreement cheap: you ask, you check, you offer the menu, you name your own weak spot. The person across the table is a teammate, not an examiner, even when they are an examiner.
The experienced do not lose the room on knowledge. They lose it on pacing.