Built a tool that turns a job posting into a timed mock interview in your IDE. What worked and what didn't
热度趋势
趋势数据积累中
百分比基于当前可用热度信号，而非评论数或独立用户人数。
一位开发者创建了一款工具，能将招聘信息转化为集成开发环境（IDE）中的定时模拟面试，填补了谷歌关闭 Interview Warmup 后留下的空白。与 LeetCode 或 HackerRank 不同，该工具提供隐藏测试、强制截止日期并跟踪每个问题所花费的时间。它通过研究公司特定的面试风格来生成原创问题，并在 IDE 中打开一个包含问题陈述、解决方案文件和示例测试的真实面试仓库。该系统通过持续集成（CI）中的变异测试门，确保所有已发布和生成问题的评分质量。
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.