The Mock Interviewer in Your Pocket
Technical and case interviews punish two things: thin fundamentals and thin reps. AI fixes the reps problem cheaply. It does not fix the fundamentals problem, and if you let it, it will hide the gap until the day you actually need to perform.
Use AI to run drills, pressure-test your reasoning, and simulate hostile follow-ups. Do not use it to think for you. The rest of this chapter is the playbook for that line.
Coding Interviews: Drill, Don't Lookup
The mistake almost everyone makes: they ask AI to solve the problem, read the solution, nod, and move on. You will retain almost none of it. Worse, you will feel prepared when you are not.
Flip it. AI is the interviewer, you are the candidate. It only helps after you have struggled.
You are a senior engineer interviewing me for a [backend / frontend / data]
role at a company like [Stripe / Meta / a Series B startup].
Give me one medium-difficulty problem. Do NOT show the solution.
After I send my approach in plain English, ask one clarifying question
and one follow-up about edge cases. Only when I say "reveal", show
an optimal solution and explain the time/space tradeoffs.
Three rules that make this actually work:
- Explain before you code. Type your approach in English first. If you cannot describe the algorithm in three sentences, you do not understand it yet.
- Code in a real editor, not the chat. Open your IDE or a blank file. Type it out. No autocomplete from AI.
- Time yourself. 25 minutes for a medium, 45 for a hard. If you blow through, that data point matters more than the right answer.
After the drill, paste your code back and ask:
Here is my solution. Don't rewrite it. Tell me:
- Where would a strict interviewer push back?
- What edge case did I miss?
- Is there a cleaner pattern I should learn for next time?
This is where AI earns its keep. It will catch the off-by-one, the unhandled null, the O(n²) you could have made O(n log n). You learn the pattern once, then drill it in a fresh problem the next day.
Product, Consulting, and Finance Cases
Cases are different beasts. The interviewer is grading your structure, not your final number. AI can stress-test the structure if you set it up right.
For a product case, prompt it like this:
You are a Senior PM at [Google / a fintech startup]. Give me a product
sense question. After my answer, push back like a real interviewer:
challenge my user segmentation, ask why I deprioritized something,
and force me to defend tradeoffs. Score me at the end on:
problem framing, user empathy, prioritization, metrics.
For consulting cases (McKinsey, BCG, Bain style), do the same with a market sizing or profitability problem. The trick: solve it on paper or a whiteboard first, then narrate your structure to AI and let it interrogate you. Reading off a screen kills the muscle you actually need on interview day.
For finance cases — DCF walkthroughs, LBO mechanics, accounting questions — AI is excellent at generating realistic technicals and grading your reasoning. Pair it with AI for Equity Analysts if you are targeting buy-side or banking and want deeper coverage of the modeling side. For corporate finance and bank-adjacent roles, AI for Bankers and Loan Officers has useful drills too.
A pattern that works across all three case types:
Run me through 5 rounds. Each round:
1. You ask one case question.
2. I respond.
3. You ask exactly one tough follow-up.
4. You give me a 2-sentence critique before the next round.
Do not be polite. I want to be uncomfortable.
Five rounds, thirty minutes, no AI doing the thinking. That is a real practice session.
The "Don't Get Dependent" Rule Set
This is the part most people skip. Read it twice.
Rule 1: No AI during the first attempt. Ever. Read the problem, sit with it, write something down. Even if it is wrong. The discomfort of not knowing is exactly the skill you are training.
Rule 2: AI critiques, it does not author. If you find yourself copying its solution into your notes verbatim, stop. Close the chat. Re-solve the problem from scratch tomorrow. If you cannot, you did not learn it — you just witnessed it.
Rule 3: One unassisted set per week, minimum. Use LeetCode, Pramp, or a friend. No AI in the loop. This is your calibration. If AI-assisted you is crushing it and unassisted you is freezing, you have a problem only real-life reps will solve.
Rule 4: Track your "first-try solve" rate. Not your "eventually got it with hints" rate. The interviewer does not care that you would have gotten it with hints.
Rule 5: Practice talking out loud. Type your reasoning into the chat as if you were saying it. Better, actually say it. Real interviews are verbal. Silent solvers freeze when forced to narrate.
Building a Two-Week Drill Plan
Cram does not work for interviews. Spacing does. A workable shape:
- Mon / Wed / Fri: one coding problem, 30 minutes, then a 10-minute AI critique pass.
- Tue / Thu: one case (product, consulting, or finance depending on your track), AI as interviewer.
- Saturday: one full unassisted mock — pen, paper, or a real platform. No AI.
- Sunday: review the week. Ask AI to summarize your weak patterns from the chat history and pick next week's focus.
Here are the problems and cases I worked through this week:
[paste topics + your performance notes]
What are the 2-3 patterns I keep missing? Build me next week's
drill plan focused only on those.
That is the loop. Solve, get critiqued, expose the gap, drill the gap, repeat.
When AI Is Wrong, And It Will Be
AI hallucinates complexity analysis. It invents library functions. It will confidently grade a wrong finance answer as correct if your structure sounded clean. Treat its feedback as a strong second opinion, not a verdict.
If something feels off — a complexity claim, a market sizing assumption, an accounting rule — verify against a textbook, the official docs, or a human who has done the job. Especially for finance and consulting, where domain conventions matter and AI smooths over the edges.
The interviewer in the room will not be impressed that ChatGPT said you were right. They will be impressed that you can defend your answer without it.