You commented INTERVIEW, so here is the full list — fifteen questions, what to prepare for each, and a four-week plan at the end.
The one-line version: DSA is not going away. It is just not the whole interview any more. The rounds being added around it test whether you can work with an AI assistant without losing track of what it wrote.
A third of the market still bans AI in the room, so this is a transition and not a finished shift. Prepare for both. The good news is that the preparation overlaps almost entirely.
1. "The AI wrote this function. What's wrong with it?"
Prep: take AI-generated code and review it line by line. Look for edge cases, Big-O claims that are wrong, and missing error handling.
2. "Solve this with the assistant, and talk me through your prompts."
Prep: practise saying what you are asking for and why, out loud, before you type it. The narration is the thing being scored.
3. "How do you know this output is correct?"
Prep: write a quick test before you trust AI code. Every time. In the interview, reaching for a test unprompted answers this question before it is asked.
4. "What would you never let an AI do without checking?"
Prep: have a real answer — auth, payments, migrations, anything that deletes data. A candidate with no limits reads as a candidate who has not shipped.
5. "Here's a codebase you've never seen. Find the bug."
Prep: clone a mid-sized open-source repo and trace one feature end to end, from the UI down to the database.
6. "This endpoint is slow. How would you find out why?"
Prep: learn to read logs, profile code, and spot an N+1 query on sight.
7. "Refactor this without breaking anything."
Prep: practise small, safe changes with tests around them. The instinct to put the test in place first is what is being watched.
8. "Why did you choose this architecture?"
Prep: for each project, write down three decisions you made and what you traded away to get them.
9. "How did you handle authentication?"
Prep: know sessions versus JWT, and the OAuth basics, well enough to explain why you picked one.
10. "What happens when 10,000 users hit your app at once?"
Prep: caching, queues, rate limits, database indexes. Know which one you would reach for first and why.
11. "What would you change if you rebuilt it?"
Prep: have an honest answer. Interviewers love self-awareness, and "nothing" is the worst thing you can say here.
12. "How does your AI agent remember context?"
Prep: know the difference between the context window, RAG, and long-term memory — and which problem each one actually solves.
13. "How would you evaluate whether your AI feature works?"
Prep: learn what evals are and build a tiny one for your own project. Almost nobody does this, which is exactly why it lands.
14. "How would you cut the cost and latency of your LLM calls?"
Prep: caching, a smaller model for the simple steps, streaming, batching.
15. "What's MCP and when would you use it?"
Prep: build one small MCP server that connects an agent to a real tool. Having built one beats any definition you can recite.
Week 1. Ship one project with an AI feature in it, and deploy it. Deployed, not local.
Week 2. Write the "three decisions and their tradeoffs" doc for that project. Practise explaining it in two minutes.
Week 3. Read and debug an open-source codebase. Do two AI-assisted practice problems out loud, narrating every prompt.
Week 4. Add evals and a load test to your project.
Do those four things and you have real answers to more than half of this list — not rehearsed ones, which is the difference the new format is built to detect.
If you want something to build for week 1, there is 25 projects to build to get hired in 2026, eight repos that make Claude Code better for the tooling side, and the gesture-controlled portfolio as a worked example of a project with decisions worth defending.
Follow @avi_vashishta29 for more like this.