Phase One: The Bug

Phase One: The Bug#

Twenty-four of the forty minutes go here, and only about six of them involve typing code. That ratio is the single most surprising thing about a strong run in this round.

0:00 - 0:04 — Reading the file#

What most candidates do. Select all, paste into the chat panel, type why is this broken?. Thirty seconds gone, and the work has started.

What it costs. The interviewer is watching to see whether you can make sense of code you did not write. That was the question. You just forwarded it to the model, in the first minute, in full view. Everything after this is coloured by it.

What to do instead. Read the file. Four minutes on twenty-five lines sounds indulgent and is not — it is the highest-leverage stretch of the entire session.

Reading is not a general skill here, it is a search for specific things. On this file:

• Where the decision happens. allow() is the entry point, but the decision is really self.tokens >= 1, and everything that matters is upstream of it in _refill(). That is the function to understand.
• What the state means. Four fields. Three are obvious. The fourth, refill_rate, is documented in the docstring as tokens per second — and that word per is the whole bug. You cannot find an arithmetic error without knowing the intended arithmetic.
• What already exists. Are there tests? Do they pass right now? A test that was already failing before you arrived is something you want to know at minute three, not minute thirty.

Narrate it, and defend the time rather than apologising for it:

"Let me read this properly before I touch anything — I want to know what it's supposed to do before I decide what it's doing wrong."