What the round looks like
You get a small repository and a task. The assistant is in the environment and you are expected to use it: one candidate describes telling it the approach and giving feedback on what came back, without typing code. The interviewer may say almost nothing. One candidate called the round confusing for that reason, and only learned from the rejection what had gone wrong.
The tasks reported are all variations on one theme. In one, an interface already had a cache and the task was to add the same to another interface, with the emphasis on testing it. In another, the task sheet asked for a read-through cache. In a third, the candidate directed Claude to write an LRU cache, then an LFU cache, then tests, then a performance measurement. A fourth described it as adding different kinds of cache to a small project.
Why candidates who finish the task still fail
The pattern across the reports is the same. The assistant produces working code quickly. The candidate accepts it. Then the interviewer asks for the overall approach, or how the files fit together, and the candidate has not read enough of the repository to answer. One wrote that the hard part was not getting it done but explaining it. Another was told, in the feedback, to understand the problem and the context more before prompting.
Our read: the round is a comprehension test with an assistant in it. The task is small on purpose. What it measures is whether you can read an unfamiliar codebase well enough to tell the assistant what to do, check what it did against the code around it, and defend the result. Forty-five minutes is enough for that only if you read first and prompt second.
What is confirmed, and what is not
- Confirmed: a 45-minute round in the virtual onsite; a project repository with a task; an AI assistant the candidate directs; a task about caching; explaining the approach and the code as the graded part. Each appears in more than one report.
- Reported once: the recruiter's description of the round as experimental and lower-weighted; Claude as the assistant; a read-through cache as the task's wording; LRU, LFU, tests and a performance measurement as the sequence; testing as the emphasis of the mirror-the-existing-cache version.
- Not known: the language of the repository, whether candidates may bring their own tools, and how the round is scored against the others.
How we would prepare
- Read a small repository in ten minutes and say how it fits together. Which module owns what, where the contracts are, what the existing cache does on a miss and on a write. Do this before the first prompt.
- Mirror an existing pattern without copying its assumptions. A cache that was right for data that never changes is wrong for data that does. Ask what the new interface's data does differently.
- Know LRU and LFU well enough to specify them. Which key goes on a tie, what happens to a key's count when it is evicted and comes back, and how you would prove the order is right with a test.
- Measure, do not assert. Have a way to count reads that reach the store and a way to time a stream of requests, so "faster" is a number the interviewer can see.
- Narrate what you accept and reject. The interviewer who says nothing is still listening. Say why the assistant's code is right, or what you asked it to change.
Nyrion's Pinterest section starts with Read-Through Cache for the Pin Service, built from these reports. A pin and board service where boards are already cached and pins are not; you add the cache, then make writes visible, then implement an LFU policy and make it fast under load. The task shape, the read-through cache, LRU and LFU, and the emphasis on reading the files are what candidates describe. The domain, the files, the invalidation stage and the exact tie rule are ours.
How this round compares with other companies' versions, and what every version grades: the AI coding interview guide.
Frequently asked, briefly answered
Where in the process is the AI coding round?
In the virtual onsite, as one of its 45-minute rounds, after a one-hour phone screen. One candidate was told by the recruiter that the round is still experimental and weighs less than the other onsite rounds.
Do I write the code myself?
Candidates report not typing the code at all: they tell the assistant the approach, read what it writes, and give feedback. One candidate who let the assistant do everything and could not explain the files afterwards failed the round.
What is the task about?
Caching, in every report so far. One version had a cache on one interface and asked for the same on another. One asked for a read-through cache per a written task. One had the candidate direct Claude to write LRU and LFU caches, tests, and a performance measurement.
Which assistant is it?
Claude is the one named in the reports. We have not seen a report of candidates bringing their own tools to this round.
What does the feedback look like?
One rejected candidate was told to understand the problem and the context more before prompting. Another said the hard part was not getting it done but explaining the whole approach and understanding the code across files.
Practice problems tagged for Pinterest
Browse the catalog and open the Pinterest section. No login required to look; sign in when you want to solve one.