Take-home, live coding, or AI screen: which preparation actually transfers

You have maybe fifteen hours before the interview. Where those hours go should depend almost entirely on which format you are sitting, and most advice online ignores the question completely.

Here is what each format actually measures, what it punishes, and what preparation moves the needle.

The take-home

What it is really scoring

Almost never raw ability to make it work. Nearly every submission works. What separates them is everything around the code: whether it is readable, whether the tests test anything, whether the README explains the decisions, and whether the scope was judged sensibly.

Reviewers are usually asking themselves one question. If this person joined and opened a pull request like this, would I be happy? That question is not about cleverness at all.

Where people lose it

• Gold-plating. Four hours of scope on a two-hour brief reads as poor judgement, not enthusiasm, and it makes the reviewer wonder how you would estimate work.
• Tests that assert nothing. A suite that only checks the happy path is worse than three tests that cover the awkward cases.
• No README. The single cheapest way to add signal. Say what you built, what you deliberately left out and why, and what you would do next.
• Ignoring the brief. If it says use whatever stack you like, that is a real choice you will be asked to defend. If it names a stack, use it.

What to prepare

Very little, technically. Instead, decide in advance what your defaults are: how you structure a project, how you handle errors, what you test first. Take-homes reward having opinions you can apply quickly rather than inventing an approach at 11pm.

The live coding round

What it is really scoring

How you behave while writing code with someone watching. That includes the code, but it also includes whether you clarify before starting, whether you narrate, whether you notice your own bug, and whether you stay functional when something does not work.

Interviewers are trying to simulate pairing with you. A candidate who solves it silently in twelve minutes can score below one who solves it in twenty while thinking out loud, because only one of those is evidence about working together.

Where people lose it

• Silence. From the outside, a quiet candidate is indistinguishable from a stuck one.
• Coding before clarifying. Two clarifying questions cost thirty seconds and prevent solving the wrong problem.
• Panicking at a bug. Everyone writes bugs live. Reading the error out loud and forming a hypothesis is the whole skill being measured.
• Refusing to start. Waiting until the perfect approach arrives reads as paralysis. A stated brute force with 'I would then improve this by' is a strong opening.

What to prepare

Practise talking and typing at once, which is genuinely awkward the first few times and much less so by the fifth. Also practise in whatever environment they will use, because losing four minutes to an unfamiliar editor is a real and avoidable cost.

The AI screen

What it is really scoring

Whether what you say holds up when it is followed up on. There is no rapport, no benefit of the doubt, and no interviewer filling in the half of your answer you left implied. What you said is what you get credit for.

It is also the round where vagueness is most expensive, because a follow-up is generated from your own wording. Loose language is not glossed over, it becomes the next question. There is a fuller breakdown in what actually happens in an AI technical screen.

Where people lose it

• Naming things without explaining them. The follow-up will ask, and there is nowhere to go.
• Long answers with no conclusion. Structure matters more here than anywhere else.
• Treating it as low stakes because no human is present. A human reads the summary and decides.
• Bad audio. A transcription error costs you an entire answer and you will never know it happened.

What to prepare

Depth on the specific technologies where your CV and the job description overlap, rehearsed out loud. Three layers on each: what it is, how it works underneath, how it fails.

What transfers and what does not

Some things do transfer across all three: leading with the answer, naming trade-offs, quantifying scale, and being explicit about where your certainty ends. That technique is worth practising regardless. The format-specific parts are not interchangeable at all.

Should you decline a take-home?

A reasonable question and a real trade-off. A two to three hour take-home from a company you want is usually worth it, particularly if you interview badly under live pressure, since it is the format most forgiving of nerves. An open-ended one with no time cap, or one that looks suspiciously like unpaid feature work, is worth pushing back on.

Asking whether there is an alternative round is normal and rarely held against you. Companies that penalise the question have told you something useful.