Take-Home Brief With the AI Rule Stated

Notes for the sender

Delete this block before sending. Everything after it is the brief.

What to edit. Replace every bracketed placeholder. Keep one of the two blocks under The AI Rule and delete the other. Then send identical wording to every candidate for the role, because two candidates answering differently worded briefs are sitting two exercises with one name.

Choosing the block. Copy the rule off the seat rather than off a policy page. Ask the people already doing the job which assistants they can open on a work machine, what they may never paste into one, and what a normal week's output owes to a model. A work sample earns its validity by approximating the real work situation ¹, so Block B belongs where the honest answer is that no third-party model may touch the work: an air-gapped system, records that cannot enter a third-party service, or work whose value is unaided recall. Everywhere else Block A samples the job and Block B samples obedience to a rule the first week reverses. Do not assume the answer either way. In a survey of 5,273 employed adults in the United States, 55% said they rarely or never used an AI chatbot at work ².

Choosing the hours. Two hours is a convention rather than a measured finding, so set the number from the reading rather than from what sounds fair. Time yourself reading one past submission against the criteria you plan to use, multiply by the finalists you expect in a month, and hold the result against the week you have. If it does not fit, the brief is too long whatever a candidate spends on it. Cutting the exercise costs coverage rather than depth: a shorter brief carries one competency, so the rest of the loop has to carry the others and the debrief stops claiming a candidate is broadly strong. Above two hours you are asking for an evening, which is not equally available to a person holding a job or a second shift, so pay for the time or cut the brief.

Why the bound sits on the deliverable. A stated time binds only the candidates who honour it, and the finished page no longer carries the hours. Unedited model output submitted through a real university examinations system drew no concern of any kind from markers in 94% of cases ⁶. The bound under The Deliverable is the part that reaches everybody, which is why it is written down.

What Block A asks for, and why those four things. The acts worth reading do not appear in a finished document. The thinking that survives assistant use shifts toward verification, integration and oversight ³, so the extra page asks for the question, the assumption, the check and the rejection. It does not ask how much the assistant helped, because self-report on that is unreliable in a measured way: sixteen experienced developers took 19% longer with AI tools on their own repositories and still believed afterwards that the tools had sped them up by 20% ⁴.

If a machine reads the submission. A person reading a submission is one thing, and an automated tool that makes or informs the decision is another, and the second carries notice duties where you hire. Illinois: Artificial intelligence with a discriminatory effect, and zip codes used as a proxy, are civil rights violations; an employer must also notify the employee that AI is used. 775 ILCS 5/2-102(L), effective January 1, 2026. New York City: A bias audit within the prior year, a public summary of its results, and notice to the candidate 10 business days before the tool is used. NYC Admin. Code 20-870 et seq., effective January 1, 2023. Where either reaches you, that notice runs on its own schedule and this brief is not it.

Freeze it. Once the first candidate has the brief, a change goes to everyone who already has it with the deadline extended, and the file records which version each submission answered. An unrecorded edit halfway through a round makes the early and the late submissions two different tests.

The task

[One paragraph naming what to produce and the situation it sits in. Name the material to work from, and the question the work has to answer.]

[If the material carries a genuine ambiguity, leave it there and say nothing about it. It is the part that separates two submissions.]

The deliverable

• [Format and bound: two pages, five slides, one function against the signature below, with a word limit where the format has one.]
• [Anything else expected back, or the word "nothing".]

Extra polish earns nothing here. A cover page, a colour palette and an appendix are not read.

If the time runs out, hand in what exists and add two lines on what would come next. That is the intended way to finish, and it is not held against you.

The time box

Spend no more than [n] hours on this, and stop at [n] hours.

The number is a cap, and nothing in this exercise rewards a longer evening. The bound that is actually read is the one on the deliverable above.

Materials and scope

Questions go to [name] at [address], answered within [n] working hours. Anything one candidate asks that changes the brief is sent to every candidate on the role.

The criteria

The submission is read against these, in these words:

[n] people read every submission against that list, and the list is fixed before the first submission arrives. The weighting and the reference answer stay unpublished.

A short conversation about the submission follows. One decision is picked out, and the question is what would have changed if that input had been different.

The AI rule

[Keep one block. Delete the other, rename the kept heading to The AI Rule, and delete this line.]

Block A. Assistant allowed, use recorded

An AI assistant is allowed on this exercise, and most of the team doing this work uses one. A free tier is fine, and the choice of tool is not read.

Send one extra page with the deliverable:

1. The questions asked before anything was generated.
2. The assumptions written down.
3. What was checked against a source outside the assistant.
4. One thing the assistant produced that was not used, and why.

That page is read alongside the deliverable. How much of the assistant was used is not read, and using none is a complete answer. Nothing disclosed on that page counts against you.

What this block shows: how a problem gets framed, what gets delegated, what gets verified, and what gets turned down. None of those four appear in a finished document, which is why the extra page exists.

What it does not show: unaided competence. Where the seat runs without these tools, this block samples the wrong work.

Block B. No assistant

Work without an AI assistant on this exercise. [One clause with the reason, taken from the seat itself: the records this role touches cannot enter a third-party service, the work is performed on a system where no such tool is available, or the value of the work is unaided recall.]

What this block shows: unaided competence, where the seat genuinely requires it.

What it does not show: that the rule was kept. Nothing in a returned document establishes which tools produced it, so this block runs on trust and on the conversation described above. No claim of that kind is made here, and none belongs in a brief.

Dates and terms

• Sent [date]. Due [date] at [time and time zone].
• [Paid or unpaid, in one line. Where it is paid, the amount and how it arrives.]
• The submission is kept for [n] months, and is read only by the people named above.
• [Adjustments: who to write to, and by when, for a change to the format or the timing.]

Limits

This is clause text for an employer's own counsel to edit. It is not advice, and it is not a complete assessment process. It applies no law to your facts.

It states no threshold, no deadline and no filing requirement of its own. Every legal statement in it is a reference resolved against a dated source, and where you hire decides which of them reach you.

Nothing here establishes which tools produced a document, and nothing here produces a number that stands for a person.