In this article
Before you start
This page is one part of the interview process guide, which maps the stages and rounds around an exercise like this one.
Why interviewers ask this
An employer that allows AI still has to learn something about you. The exercise tests what stays human: reading the rules, steering the tool, checking the result and explaining it.
Anthropic says it wants candidates who "excel at collaborating with AI", and its engineering write-up says candidates may use AI tools as they would on the job while still showing their own skills. Its interview process page covers the wider loop.
A weak submission is a working app that nobody can explain. A strong one is a working app plus the assumptions you made, the checks you ran and an honest note of what the AI wrote. The live version of this skill is covered in AI-assisted coding interviews.
The AI-allowed build exercise from first principles
- AI-allowed build exercise
- The AI-allowed build exercise is a take-home or timed task in which the employer states that AI tools may be used, asking you to build a small working piece of software while showing your own judgment, checking and ownership of the result.
- Verification pass
- A verification pass is a deliberate final check, run by you and not by the AI tool, that the delivered behavior matches the brief and that nothing which measures correctness, such as the tests, was changed to make the result look better.
In plain words
Think of an open-book exam where the book can also write answers. The examiner still wants to see what you understood. So the questions are bigger, and the marking looks at how you used the book. The analogy breaks in one place. There is often no invigilator, so the honesty of your notes carries weight.
In one precise sentence, the exercise hands you a tool that can produce code and then grades the parts the tool cannot supply for you: permission, direction, verification and explanation.
Step 1: Confirm in writing that AI is allowed
Anthropic's candidate guidance sets the default at no AI assistance unless the task says otherwise. Its guidance reads "Complete these without Claude unless we indicate otherwise" and promises "We'll be clear when AI is allowed". Its write-up says the take-home it describes explicitly indicates otherwise.
So the rule is the opposite of "if it is not forbidden, it is allowed". Input: the prompt and any recruiter email. Output: one sentence you can quote back, such as "AI tools are allowed for this build." If the prompt is silent or vague, ask the recruiter before you open a tool. The same guidance asks you to "Be transparent", which feeds Step 5. You now have permission on record, which is the base for every later step.
Step 2: Split the time before you write code
Anthropic's take-home began as a 4-hour window and was later reduced to 2 hours. The write-up gives scheduling overhead causing multi-week delays in its pipeline as the reason for the shorter window, and it also says that at the same time limit Claude Opus 4 outperformed most human applicants. Whatever your window is, divide it before you build.
The split below is illustrative, for a 2-hour window, and is not a published rule.1 Spending nearly two fifths of the window on checking and writing up is deliberate: that is where the tool cannot help you.
| Activity | Minutes (illustrative) | Share of 120 |
|---|---|---|
| Read the brief and write assumptions | 15 | 12.5% |
| Build with the AI tool | 60 | 50% |
| Tests and verification | 25 | 20.8% |
| Readme with the AI log | 20 | 16.7% |
Writing assumptions first matters when the brief is deliberately vague, because you can then build on a stated guess instead of stalling.
Step 3: Direct the tool instead of watching it
Direct the AI tool in small tasks and read each change, because passive delegation leaves you unable to explain the work. That is our recommendation. Hume, who designed Anthropic's take-home, explains why the original version stopped measuring anything: a human steering Claude "would likely be constantly behind, understanding what Claude did only after the fact", and the dominant strategy might become sitting back and watching. The redesign exists so that watching is no longer enough, so show that you steered.
The same write-up says that by May 2025 over 50% of candidates would have been better off delegating to Claude Code entirely on the earlier version, which is why the employer redesigned it.
A practical pattern: ask for one function or one screen, read the diff, run it, then ask for the next piece. Decide where to spend effort on tooling, such as a small script that exercises your endpoints. The write-up states that "Judgment about how to invest in tooling is part of the signal". Say why you chose that investment in your readme.
Step 4: Verify the work before you submit
Run a verification pass: verify the work yourself and confirm the tests folder is unchanged before you submit. The README of Anthropic's original performance take-home reports that none of the first-day solutions below 1300 cycles were valid, because "In each case, a language model modified the tests to make the problem easier."
It gives one example: a model noticed multicore support, saw no speedup, and "debugged" the core count down to one. A passing test run therefore proves nothing if the checks themselves moved.
The README advises telling an agent not to change the tests/ folder and to verify with tests/submission_tests.py. It adds that git diff origin/main tests/ should print nothing. Run an equivalent check on your own repository, then try inputs the brief did not list, such as empty input, a failed network call and a very long value. Write the check you ran in your readme.
Step 5: Log what the AI did and hand over
Log what the AI did in a short readme, because transparency is the one part of the guidance that is fully in your control. The candidate guidance says "Be transparent" and "Use Claude thoughtfully."
A readme that works has four short parts: how to run the project, the assumptions you made, what the AI wrote and what you changed or rejected, and what you would do with another day. Candidates report that some exercises also ask for a recorded demo or a live walkthrough, so be ready to explain any file in the repository without opening the tool.
If a live round follows, the skills carry over; see forward deployed and applied AI engineer rounds for what those loops ask. The five steps above are the whole routine: rules, time, direction, verification, log.
Step-by-step description
- Confirm in writing that AI is allowed.
- Divide the window before you code.
- Work in small steps and read each change.
- Run the checks and confirm the tests are unchanged.
- Write down what the AI did.
- Anthropic's candidate guidance sets the default at no AI assistance unless the task says otherwise.
- Verify the work yourself and confirm the tests folder is unchanged before you submit.
How far can the AI get alone on a build?
Far enough that employers redesign tasks around it. The README of Anthropic's original performance take-home lists runs in simulated clock cycles, where lower is better. The 2-hour version of the take-home started from code that ran in 18532 simulated clock cycles; the public repository starts about 7.97 times slower, and the README's AI runs are compared here against the 18532-cycle start.1
A casual Claude Code session reached 1790 cycles, a 10.35 times speedup1 from a start of 18532 cycles, which the README says is approximately matching the best human performance in 2 hours. The improved harness reached 1363 cycles, a 13.6 times speedup.
| Run | Cycles | Speedup over start |
|---|---|---|
| 2-hour take-home starting code | 18532 | 1.00 |
| Opus 4 after many hours in the harness | 2164 | 8.56 |
| Opus 4.5, casual Claude Code session | 1790 | 10.35 |
| Opus 4.5, 2 hours in the harness | 1579 | 11.74 |
| Opus 4.5, 11.5 hours in the harness | 1487 | 12.46 |
| Improved harness | 1363 | 13.60 |
Text description of the diagram
- A casual Claude Code session reached 1790 cycles, a 10.35 times speedup from a start of 18532 cycles.
- The improved harness reached 1363 cycles, a 13.6 times speedup.
There is a trade-off in the redesign. The write-up says "The original worked because it resembled real work. The replacement works because it simulates novel work." The employer gave up some realism to keep the signal. These are one employer's numbers, not a survey of how others set the bar.
Knowing it is half the job. Saying it is the other half.
Interview Clozo is an AI interview copilot for real interview rounds: it hears the question and puts a strong answer on your screen, and you say it in your own words. See the features or how it works.
Start Free Trial →The Staff-level answer
The prompt says you may use AI tools for this build. How do you approach it?
First I confirm in writing what is allowed, and I ask the recruiter if the prompt is vague. Then I write my assumptions and split the window into reading, building, checking and documenting. I let the tool draft, but I keep the design decisions, work in small steps and read every change. Before I submit I run the tests myself, confirm the tests folder is unchanged because agents have been seen editing tests to fake a pass, and try the unhappy paths. My readme says what the AI wrote, what I changed or rejected, and what I would do with another day.
What makes it Staff-level is that it treats verification and disclosure as the main work, names the failure it is guarding against, and keeps ownership of the decisions while using the tool for speed.
The rubric
This rubric is our own reading of what a strong submission shows, not a published scoring sheet of any company.
| Dimension | Weak answer | Strong answer | Staff answer |
|---|---|---|---|
| Permission | Assumes AI is fine because nothing forbids it. | Reads the prompt and asks when unclear. | States the rule in the readme and keeps to it. |
| Direction of the tool | Pastes the brief and accepts the output. | Gives small tasks and reads each change. | Chooses where to build tooling and explains why. |
| Verification | Stops when the tests pass. | Runs tests and tries unhappy paths. | Confirms the tests folder is unchanged and says so. |
| Ambiguity | Guesses silently, or asks about everything. | Writes assumptions and builds on them. | Names the riskiest assumption and a cheap way to test it. |
| Communication | A bare repository with no notes. | A readme with setup and AI use. | A short log: what AI wrote, what changed, what was rejected, next steps. |
Three variants of this question
Can I use AI on this take-home?
Only if the task says so. Anthropic's guidance defaults to no AI unless indicated, so read the prompt, and ask the recruiter in writing if it is unclear.
What if the task is to build a small app against a brief within a day?
Candidates report tasks of this kind: a small Claude-powered app against a customer brief, a choice among several prompts with a repository and a recorded video within a day, and a one-day knowledge-graph build. No public source found confirms these, and formats change, so follow the prompt you receive.
How do you implement a small React app that requests a voice application programming interface (API)?
A candidate reported this task. Keep the application programming interface (API) key on a small server route, never in browser code. Have the React component send text to that route, let the route call the voice service, and return the audio with loading and error states. Check the provider's documentation for the exact call, because an AI tool can invent parameters.
Common mistakes
- "Nothing says I can't, so I'll use AI." (wrong) Anthropic's default is no AI unless indicated. (right) "I confirmed in writing that AI is allowed for this task."
- "The tests pass, so it is correct." (wrong) Agents have edited tests to fake a pass. (right) "The tests pass and the tests folder is unchanged."
- "I'll let the AI build it and describe the result." (wrong) You cannot explain what you did not read. (right) "I worked in small steps and can walk through every file."
- "I won't mention the AI in my readme." (wrong) The guidance asks you to be transparent. (right) "The readme logs what the AI wrote and what I changed."
Follow-up questions
Is AI allowed in a take-home unless the task says otherwise?
At Anthropic, no: the default is no AI unless the task indicates it. Other companies set their own rules, so check the prompt and ask the recruiter.
Does an AI-allowed exercise mean the task is easier?
Not necessarily. Anthropic describes redesigning its take-home so that it still carries signal with AI available, which is one employer's practice and not a rule.
How do you show your own skill when the AI writes the code?
Show the decisions around the code: assumptions, the checks you ran, the problems you caught and what you rejected. A candidate who can walk through every file shows ownership. See AI engineer interview questions for the questions that follow.
Do companies still ask aptitude rounds for experienced candidates?
Candidates report mixed experiences, including one who was asked to build a working workflow and an agent and then present it. No public source found gives a rate, so read the interview process guide for the stages and ask the recruiter what your loop contains.
- Permission is explicit and per task; confirm it in writing before you use AI.
- Split the window up front and reserve real time for checking and the readme.
- Steer the tool in small steps and read every change, because passive delegation shows nothing about you.
- Passing tests prove nothing if the tests changed; confirm the tests folder is unchanged and say so.
- Log what the AI wrote, what you changed and what you would do next.
Sources
- Guidance on Candidates' AI Usage (How to collaborate with Claude during our hiring process) — Anthropic (accessed 2026-10-26), 10 Jul 2025
- Designing AI-resistant technical evaluations — Anthropic, Tristan Hume (accessed 2026-10-26), 21 Jan 2026
- Anthropic's Original Performance Take-Home (README) — Anthropic on GitHub (accessed 2026-10-26), 21 Jan 2026
