AI-Assisted Coding Round (HackerRank + Claude)

An onsite/tech-screen round run inside HackerRank with an embedded AI assistant. Candidates are given a problem — most often a rule parser that evaluates accept/block conditions over a list of transactions — and must solve it by driving the AI rather than coding by hand. How you prompt, validate, and reason about the AI's output is the graded signal. Requirements The round runs inside HackerRank with an embedded AI chat assistant (reported as Claude). A detailed README ships with the task; read it first. Canonical task: write a rule parser. You are given a list of transactions and a list of rules; each rule names an action (accept / block) followed by an if clause. Match keywords in a transaction against keywords in the if clause (plain string equality — no fuzzy matching) and apply the rule's action. Part 2 extends the if clause grammar to support a boolean operator (e.g. and / or over multiple conditions). Only ~30 minutes is actual coding time; in Java loops candidates report relying on the AI for roughly 80% of the code to finish in time. Notes This format began appearing in loops around April 2026 and now shows up in many VO loops, replacing one traditional coding round. The recruiter sometimes only warns you a few days before. The task itself is easy — easier than the phone-screen / OA coding problems — but the rubric is the hard part and is openly murky. Candidates report contradictory interviewer signals: lean too hard on the AI and you get "why are you so confident in this code?"; write your own skeleton and the interviewer asks "why aren't you using the AI?". Naming and test cases get nitpicked. An approach that earned a "solid good" grade: have the AI read the README and propose a plan, review the plan, have it write tests, then run the tests — drive it like a junior engineer rather than asking for a one-shot answer. The AI may over-engineer or silently drop a condition; reviewing and correcting its output is where the signal lives. A recent L2 loop reinforces that speed and active delegation both matter: methodical step-by-step verification can still be judged too slow, with feedback to use the assistant more aggressively. Keep review cycles short and let the AI implement routine edits while you own the plan and validation. Preparation Practice driving an AI coding assistant on small problems: state the problem, ask for a plan, ask for tests, iterate on edge cases, run the tests. Rehearse narrating why you accept or reject each AI suggestion — interviewers interrupt to probe your judgment on code you didn't type. Build the habit of validating AI output against tests rather than trusting the first response, and be ready to write a clean skeleton yourself if asked.