Fix the shape of the day and let the topics fall into it. Ninety minutes daily in four blocks: 25 minutes on one pattern, two timed 25-minute problems back to back, 15 minutes writing a post-mortem. Twenty-one problems are assigned across days 1 to 6, day 7 is one 45-minute mock plus a machine check, and each day ends with a pass or fail checkpoint, so on day 3 you learn whether to cut scope instead of finding out mid-call.
The 90 minutes, in four blocks
A topic list is not a plan, because it never tells you when to stop. The blocks below are the plan: same order, same clock, seven days. On a 60-minute day, drop the second timed problem, never the post-mortem.
25 min, pattern review
One pattern only. Write the template from memory first, on paper, then read that day's worked solutions and fix what you got wrong. The template is what you are memorising: the four lines of a sliding window, the while loop of a binary search on the answer, the visited set of a graph traversal.
25 min, timed problem one
Timer visible. Say the brute force out loud in the first two minutes, name the pattern, then code. At 25 minutes you stop whether or not it works. Stopping mid-solution is not failure, it is the data the checkpoint needs.
25 min, timed problem two
Immediately, no break. A real screen is two problems in one hour with a tired brain, and the second one is where people fall apart.
15 min, written post-mortem
Five lines per problem, in one running file you re-read each morning during block one. This is the block people skip, and the only one that compounds.
The 21 problems, assigned by day
Nine are read as worked solutions in block one, twelve are solved under the timer. Do not substitute favourites: each day's timed pair uses the pattern you reviewed twenty-five minutes earlier.
| Day | Pattern block, read these | Timed pair, 25 min each | Pass checkpoint |
|---|---|---|---|
| 1 | Hash maps and two pointers. Read: Two Sum, Group Anagrams | Valid Palindrome, 3Sum | Both reached working code inside the timer, brute force counts |
| 2 | Sliding window and prefix products. Read: Best Time to Buy and Sell Stock, Minimum Window Substring | Longest Substring Without Repeating Characters, Product of Array Except Self | At least one finished optimal and passed the sample inputs |
| 3 | Binary search and intervals. Read: Search in Rotated Sorted Array, Merge Intervals | Koko Eating Bananas, Insert Interval | Cut-scope test. 4 of the 6 timed problems so far ended in working code |
| 4 | Trees, BFS and DFS. Read: Binary Tree Level Order Traversal | Validate Binary Search Tree, Lowest Common Ancestor of a Binary Tree | Both traversal templates written from memory in under 4 minutes |
| 5 | Graphs and heaps. Read: Number of Islands | Course Schedule, Top K Frequent Elements | Adjacency list built without looking at yesterday's code |
| 6 | One-dimensional dynamic programming. Read: House Robber | Coin Change, Longest Increasing Subsequence | Recurrence and base case stated aloud in under 60 seconds |
| 7 | Re-read the post-mortem file only. No new problems | One 45-minute mock, then the machine check below | Narrated continuously, never silent for longer than 20 seconds |
The day 3 checkpoint, and exactly what to cut
By the end of day 3 you have run six timed problems. Count how many ended in code that ran and passed the given examples inside 25 minutes. Four or more is a pass, carry on unchanged. Three or fewer is a fail, and a fail means the plan is too wide for the week you have, not that you should push harder.
On a fail, cut in this order and stop as soon as the day fits back inside 90 minutes:
- Drop day 5 entirely. Graphs and heaps are the least likely of the six to appear in a first screen and the most expensive to learn cold. Repeat day 2 instead.
- Drop the second timed problem on days 4 and 6, and spend those 25 minutes re-solving something you already failed. A second attempt builds recall faster than a first attempt at a new problem.
- Shrink day 6 to House Robber and Climbing Stairs only. Coin Change and Longest Increasing Subsequence are worth real time later, not now.
What you never cut: the timer, and the post-mortem. A week of untimed practice teaches you to solve problems in fifty minutes, which is a skill nobody is testing.
The 15-minute post-mortem, five lines
One running file, newest at the top. Concrete beats tidy.
- Trigger. The phrase in the problem that should have fired the pattern. "Longest substring with a constraint."
- Technique. The one-line template. "Expand right, shrink left while the window is invalid, record the max."
- The stall. The exact minute you stopped moving. "Minute 11, could not decide what to store in the map."
- The miss. The bug the examples caught. "Left pointer jumped backwards because I did not take the max of the two indices."
- The sentence. The finished solution explained in one breath. This is the line you rehearse aloud.
By day 6 the file runs to about sixty lines, and it is the only revision material you need. For extra problems in block one, Code Copilot captures a prompt with Ctrl+Alt+S and returns the approach, a dry run and the Big-O. Read the intuition before the code, or the block quietly turns into copying.
Day 7: the 45-minute mock and the machine check
Run the mock at the hour your interview is scheduled, with a real timer, out loud, alone if nobody can sit in. Pick two unseen problems from days 1 to 4 and keep talking through the stuck parts, because narration is scored as heavily as code.
Then the machine check, twenty minutes that prevent the failure mode with nothing to do with algorithms:
- Join a test call on the platform named in the invite, share your screen, and inspect the shared view on a second device. Banners, open tabs and file names all show.
- Turn on do not disturb, quit chat clients, close every window you will not use.
- Open the editor you plan to use and run a hello world. Do not discover a broken interpreter at minute two.
- Test microphone and camera inside the platform itself, not in system settings.
- Plug the laptop in, and put water within reach. Both matter more than one extra problem.
If a behavioural round is stapled on, write out the four or five answers you know are coming. The company question banks and the Q&A generator build a role-specific list in minutes.
Where a one-week plan honestly fails
Ten hours cannot buy fluency you do not already have. If you are still looking up how to iterate a dictionary, or have never written a recursive tree traversal, this plan hands you six failed checkpoints and a worse mood. Spend the week on primitives and easy problems instead, and walk in aiming to reason clearly and take hints well, which interviewers do reward.
A week is also wrong for a four-hour on-site loop, for senior system design, and for take-homes, graded on tests and structure rather than speed. Tooling has limits too. OfferPilot AI runs on Windows 10 and 11 only, so on a Mac or Chromebook the day 7 machine check applies and the software does not. Live assistance is out of bounds in a proctored assessment or a lockdown browser, and the overlay's exclusion from screen capture hides it from screen sharing and screen recording only. It does not defeat a webcam, a lockdown browser, a proctor in the room, or a phone pointed at your desk. Preparation is the part that always works.
If you have three days, not seven
Run days 1, 2 and 4 in order and skip the rest. Hash maps, two pointers, sliding window and tree traversal are the densest four patterns in the set, the checkpoints apply unchanged, and you arrive with about thirty lines of post-mortem. Far better than skimming twelve topics once.
Start day 1 tonight. Keep the timer honest and the post-mortem written, and cut early rather than late. For a study partner in the read-first block, the free tier gives 5 minutes a day with no card, and paid plans are on the pricing page.