Built a tool that turns a job posting into a timed mock interview in your IDE. What worked and what didn't
Heat trend
Collecting trend data
The percentage is based on available heat signal, not comment count or independent people.
A developer created a tool that transforms job postings into timed mock interviews within an IDE, addressing the gap left after Google shut down Interview Warmup. Unlike LeetCode or HackerRank, this tool provides hidden tests, enforces deadlines, and tracks time spent per question. It researches company-specific interview styles to generate original questions, opening a real interview repository with problem statements, solution files, and sample tests. The system ensures grading quality through a mutation gate in CI for all shipped and generated questions.
Problem: timed coding screens are their own skill. Four questions share seventy minutes, triage kills more sittings than algorithms do, and practicing untimed on a problem site trains none of that. Google shut Interview Warmup down in April and nothing really replaced it.
Closest existing things, and what's different: LeetCode or HackerRank with a self-imposed timer gets you reps, but there are no hidden tests, no hard stop, and no accounting of where the minutes went. Human mock-interview platforms are realistic but scheduled and usually paid. "Interview me" prompts in a chat window have no real files and grade by vibes. What I built instead: paste a job posting (or name a company), an agent researches what that company's screen actually looks like, writes an original question in that shape, and about 2 minutes later your editor opens on a real interview repo: problem statement, solution file, sample tests, clock running. Hidden tests grade submits with partial credit, a script enforces the deadline, and the report afterward shows time spent per question. For evidence of grading quality: all 22 shipped questions and 4 projects pass a mutation gate in CI (reference solution passes, untouched starter fails, every deliberately-wrong solution is caught by at least one hidden test), and generated questions pass the same gate before the clock starts.
What I did and learned: built almost entirely with Claude Code, including the Python engine, with me reviewing everything that grades people. What worked: model owns the words, script owns the numbers. The clock is timestamp math in a state file and late submissions die on an exit code, because the model's own sense of elapsed time is confidently wrong. Exit codes as the agent's API made behavior predictable. What didn't work at first: trusting the model's test suites. A rolling-median question sat behind twenty hidden tests while the classic wrong solution passed all of them, because every fixture accidentally dodged the bug. That failure became the mutation gate above.