CalibrdReady. Confident. Hired.← All guides

Interview prep · interview formats

Take-home assignments: what reviewers grade, and how long to spend

A take-home assignment is a work sample you do on your own time and then defend in a follow-up interview. The companies that publish their take-homes agree on what they grade: a working core before extras, code that runs, tests, a README that says what you assumed and what you cut, and whether you can explain every line. They disagree on how much effort to put in, and in 2026 on whether you may use AI. This page sets out what each one publishes, so you can tell which kind you have.

Updated 13 October 2026 · from company handbooks, blogs and job ads read that day

The short version

  • Make it run on a clean machine first. Reviewers say a great many submissions fail right there.
  • Build a working core, then write down what you cut and why. That list is your follow-up interview.
  • Read the brief for the AI rule. In 2026 it is banned, allowed with disclosure, or expected, depending on the company.
  • When the time limit is vague and the deadline is a week away, ask how they compare submissions.
  • Never run a stranger's repository blind: fake take-homes have carried malware.

What companies publish about their take-homes

Each of these is from the company's own handbook, blog or job ads. Where we have a full guide to the company, its name links to it.

PostHog

SuperDay: a paid full day

$1,000. A deliberately broad task you won't finish; a working core first, no gold-plating, and they review your git history. AI expected on the project, banned in the debugging session.

GitHub

A timed task submitted as a pull request

A private repo from its Interview-bot, automated tests and a rubric, reviewed anonymously. The timer exists so as not to favour people with more free time (GitHub's 2022 post).

GitLab

Review a merge request, then defend it

Up to an hour reviewing, then a 90-minute call improving it live. AI tools allowed if you say which ones.

Basecamp

A 3 to 5 hour task, with a week to do it

Given to 1 to 3% of applicants. DHH wrote that the hire's submission was borderline gold plated, and that smaller ones could just as well have won.

Linear

A paid work trial of 2 to 5 days

For every role, at a daily rate. Linear warns against hiding weak ideas behind complexity.

Automattic

A paid trial

$25 an hour, judged on the work, communication and documentation.

Firecrawl

A paid trial of about 40 hours

At a contractor rate, in 28 of 32 open roles on 13 October 2026.

Sierra

A 2-hour build with your own AI tools

Supervised on site. Criteria shared in advance, including that cutting scope is fine.

Anthropic

A timed take-home, 2 hours

No AI unless the brief allows it. Cut from four hours after Claude matched its best candidates; built so strong candidates don't finish.

Surge AI

A published written exercise

Comes with its own lists of what they look for and common pitfalls, such as overly complex workflows.

Cognition

Build your own coding agent in about 8 hours

Described by its CEO; measured on how far you get.

DoorDash

A 60-minute session on your own machine

A take-home moved into the room: your editor, your AI agent, scored on how you get oriented and check the AI.

What reviewers grade

The companies above and the reviewers who write about it agree on these:

How long to spend

This is where reviewers genuinely disagree. DHH wrote that Basecamp's hire submitted something borderline gold plated, and that smaller submissions could just as well have won; PostHog tells candidates not to gold-plate. But candidates report being called sloppy and incomplete after keeping to a two-hour limit while others did more, and one product management rubric scores who worked the hardest. A reviewer put the problem plainly: some candidates take the time limit seriously and others go far beyond it.

The evidence points to one rule. When the company enforces the time (GitHub's timer, Anthropic's two hours, Bending Spoons' proctoring) or publishes what it grades (PostHog, Sierra), keep to it and write down what you cut. When the limit is a suggestion and the deadline is a week away, assume you are compared on completeness and ask the recruiter how they compare submissions. Candidates have turned down unpaid tasks of 20 hours or more, and some who put in 30 hours report being ghosted anyway.

AI in take-homes in 2026

If the brief says nothing, ask before you start. When AI is allowed, keep a short note of what you asked it for and what you changed; it answers the follow-up question before it is asked.

The follow-up interview

Almost every take-home ends in a call about it. The questions companies and reviewers name: why this architecture, why this feature first, what you would do with more time, what shortcuts you took and why, what you found hard, and how you would add one more feature. GitLab and Sierra ask you to change the code live. Your README's “what I cut” and “what next” sections are the script; rehearse extending your code by one feature out loud before the call.

Other roles

Check before you clone

Fake recruiters, often for crypto roles, have sent take-homes whose repositories run code the moment you open them in your editor. Before you open one: check the company and the recruiter exist, read the setup files and any editor tasks or install scripts, and run it in a separate environment if anything looks automatic.

How common take-homes are

In CoderPad's 2025 survey of more than 5,000 developers and recruiters, 21% of teams used take-homes, and 12% of developers had refused one, most often over the time it would take. In its 2026 survey, 33% of developers said take-homes best reflect the job, against 12% of recruiters.

Questions

What is a take-home assignment?

A take-home assignment is a work sample you complete on your own time, usually after a screen, and then defend in a follow-up interview. Published examples run from one hour, GitLab's merge request review, to a paid five-day trial at Linear.

How long should a take-home assignment take?

As long as the company says, if it publishes a time limit and grades your decisions: GitHub uses a timer, Anthropic cut its performance take-home from four hours to two, and Basecamp's run three to five hours. When the brief is vague and the deadline is a week away, ask how long they expect and how they compare submissions, because reviewers disagree on effort.

Can I use AI on a take-home assignment?

Only if the brief says so. In 2026 the rules split three ways: Anthropic bans AI unless told otherwise, GitLab allows it if you say which tools you used, and PostHog expects it but asks you to explain every line. Read the brief, and ask when it says nothing.

What do reviewers look for in a take-home?

That it runs on a clean machine, a working core before extras, readable commits, tests, edge cases handled, and a README with your assumptions, how to run it, what you cut and what you would do next. And that you can explain and defend every line in the follow-up, especially any written with AI.

Are take-home assignments paid?

Some are, and more are becoming paid work trials. PostHog pays $1,000 for its one-day SuperDay, Automattic $25 an hour, Linear a daily rate for two to five days, and Firecrawl a contractor rate for about 40 hours. Most unpaid take-homes are meant to take a few hours.

Before the follow-up

Rehearse defending your take-home out loud

Paste the posting you are applying to. Calibrd predicts the rounds and the questions for that role, and you can practise explaining your decisions out loud in a mock interview before the real call.

Scan your posting →Add to Chrome, it reads the posting for you →

Free to start · No card · Encrypted at rest, never used to train AI, remove anytime

Sources

Take-Home Assignments: What Reviewers Grade — Calibrd