Sit down 55 minutes before the call and be finished 15 minutes before it. In that window you re-solve two problems you already know cold, narrate one problem out loud for 90 seconds so your talk-while-coding voice is already running, and skim your pattern trigger list for three minutes. Then you stage the laptop and leave the desk. Nothing new, nothing unsolved, no browser. The warm-up moves the cold start off the interview. It does not teach you anything.
The 40 minutes, counted backwards from the call
Every block is timed from T-0, the moment the call starts. Count backwards, not forwards: a routine that begins "about an hour before" ends up still running at T-2.
| Clock | What you do | Muscle it warms |
|---|---|---|
| T-55 | Plain editor open, timer visible, phone in another room. | None. This is the fence around the block. |
| T-55 to T-43 | Problem one, from a blank file, 12-minute hard cap. | Recall |
| T-43 to T-33 | Problem two, 10 minutes, typed out rather than sketched. | Syntax |
| T-33 to T-30 | 90 seconds narrating aloud into your phone, then 90 seconds playing it back. | Speech |
| T-30 to T-27 | Read the pattern trigger list. Read it, do not edit it. | Retrieval cues |
| T-27 to T-20 | Stage the machine, using the checklist below. | None, and that is the point. |
| T-20 to T-15 | Slack for the thing that goes wrong, because something usually does. | None |
| T-15 | Hard stop. Stand up, walk away, water, back at T-3. | None |
Shift the grid for an afternoon slot but keep the shape. The two blocks people delete first, the narration and the hard stop, do most of the work.
The three problems on the bench, and why each is different
Use the same three every time. Familiarity is the feature: you are testing whether retrieval fires this morning, not discovering anything. You have solved all three at least twice before today.
Number of Islands, for recall
Blank file, no template, no imports pre-typed. It makes you rebuild a whole traversal skeleton from memory: bounds check, visited marking, the four neighbour offsets, and the outer loop that starts a flood from every unvisited cell. If your hands stall at "how was I marking visited again", that stall was coming anyway, and it is far cheaper at T-50 than at minute four of a shared screen. Twelve minutes, then stop whether or not it runs.
Top K Frequent Elements, for syntax
The algorithm is not the point. It forces the two idioms everyone half-remembers: building a frequency map, then ordering by value with a custom comparator. In Python that is Counter plus heapq.nlargest. In Java it is HashMap.merge and a PriorityQueue with an explicit comparator. In JavaScript it is a Map and a sort over its entries. Nobody types a comparator confidently at minute two of an interview unless they typed one that morning.
Merge Intervals, for speech
Do not type this one. Say it. Every step is a decision you must defend rather than a line you must remember: sort by start, and be ready to say why start and not end; one pass holding a current interval; extend when the next start falls inside the current end; otherwise push and reset. Three defensible decisions in 90 seconds is about the density of a real explanation, which is why it warms what typing does not. Narrating one of the two you just solved works too. The voice is what matters, not the problem.
Substitute your own freely, as long as the categories survive. Recall means a structure you rebuild from nothing. Syntax means a problem that drags your language's awkward idioms into your fingers. Speech means a problem whose correctness rests on an argument. If you do not have three you know that well, pick them from a list of common interview questions this week.
The 90-second walkthrough, second by second
Say it aloud, at conversational volume, into your phone's voice recorder. Rehearsing silently warms nothing: what seizes up in interviews is the mouth, not the brain.
0:00 to 0:15, restate the problem
In your own words, including the input type and what you return. This is the sentence that buys thinking time in the real round, so it should be automatic by T-0.
0:15 to 0:35, walk one small example
Three or four elements, spoken as concrete values rather than variables. Saying real numbers out loud is a separate skill, and the one that most often deserts people.
0:35 to 0:55, the brute force and why it fails
State it, give its complexity, then name the specific waste. "It recomputes the same overlap for every pair" is an answer. "It is too slow" is not.
0:55 to 1:20, the real approach
Invariant first, loop second. Practise the connective tissue: "the invariant is that the output stays sorted by start, so I only ever compare against the last interval I pushed."
1:20 to 1:30, complexity and one test
Time, space, and the edge case you would check next. Ending on a test rather than trailing into silence is a habit worth owning.
Then play it back, once
Count the gaps longer than three seconds and notice where they sit. The playback is what changes the habit, and it is the part everyone skips. One pass, no re-recording.
The three-minute trigger skim
Not a list of algorithms. A list of trigger words mapped to your first move, about eight lines, in your own handwriting. Reading it primes recognition, and recognition is what fails when you are cold.
- "k largest" or "k most frequent" goes to a heap of size k, or bucket by count.
- "sorted array, find a pair" goes to two pointers from both ends.
- "longest substring with a constraint" goes to a sliding window, shrunk from the left while the constraint is broken.
- "next greater" or "next smaller" goes to a monotonic stack.
- "grid, connected regions" goes to a flood fill that marks cells as it visits.
- "intervals" goes to sorting by start before deciding anything else.
- "how many ways" or "can I reach" goes to plain recursion first, memoise second.
- "shortest path, unweighted" goes to BFS with a queue, never DFS.
Three minutes means reading, not rewriting. Any line you cannot expand in ten seconds is a real gap, and you are not closing it today. If the round has a behavioural half, this is when you re-read your company notes rather than build them. A company question bank is a Tuesday task; at T-30 you only re-read what exists.
What "laptop staged" means at T-15
Seven minutes, once, in this order. The aim is that at T-3 you press exactly one thing.
- Call link open in the exact window you intend to share, client already updated if it wants to be.
- Coding pad open, language selected, one line typed and run to prove it accepts input.
- Everything else closed. Not minimised, closed, so nothing can surface mid-share.
- Notifications off at the operating system, phone silent rather than vibrating.
- The monitor decision made now: which display does the pad open on.
- Charger in, water in reach, notes face down, resume in one minimised tab.
Then leave. The last quarter hour is the highest-variance stretch of the day: anything you start there you cannot finish, and an unfinished problem at T-3 is the memory you carry into the first prompt.
Where this routine will not help
It teaches nothing. Forty minutes cannot close a pattern gap, and if you have never built a monotonic stack then the trigger line for one is a label, not a skill. All it removes is the cold start: the first problem of any session is almost always the slowest, and this moves that cost off the interview.
For a minority it makes things worse. If both problems go badly you enter the call having just failed twice, at the worst possible moment. Test the routine on a mock before trusting it on something that counts. If re-solving reliably rattles you, cut the code and keep only the narration and the skim, which raise arousal far less.
The morning is also the wrong time to install new software, ours included. OfferPilot AI's Code Copilot capture flow is quick once it is habit, Ctrl+Alt+S, drag a box, about three seconds to scroll, drag again, Enter to send, and clumsy the first time you try it under pressure. Learn it in a mock this week or leave it closed today. Two constraints, plainly: it runs on Windows 10 and 11 only, with no macOS, Linux, mobile or browser-extension build. And the overlay's exclusion from screen capture happens at the operating-system level, so it stays out of Zoom, Teams, Meet, OBS and screenshots. That is a fact about screen sharing and nothing more. It does not affect a webcam, a lockdown browser, a proctor watching in the room, or a phone camera pointed at your desk. Where an assessment forbids outside assistance, the honest window is the 40 minutes before the call, not the call itself.
What the routine buys is narrow and worth having. Your fingers have already produced a traversal today, your language's awkward idiom is loaded, your voice has already explained something technical to a room, and your machine is not a variable. That is the whole return, and it is enough.