Choose the assessment format that reveals the work your engineer will actually do, then make it as small as possible. A polished take-home, a fluent live session and a successful paid trial can each provide useful evidence. None answers every hiring question on its own.
Start with the decision you still need to make. Perhaps you know the candidate can build a service but have not seen how they investigate failures. Perhaps their technical depth is clear and you need to understand how they work with another engineer. Write that gap down before adding another stage.
This is Corellia Talent’s assessment framework for startup teams. Use it alongside a clear founding engineer brief, so the exercise measures the responsibilities you are hiring for.
Start with one work question
An assessment should answer a specific question about the role. “Can this person code?” is too broad to guide a useful exercise or a consistent decision.
For a product engineer, the question could be whether they can turn an ambiguous customer request into a small, testable change. For a platform engineer, it could be how they diagnose an unreliable service and decide what to measure first. For a senior technical hire, it might concern a migration decision with competing constraints.
Work-sample guidance from the US Office of Personnel Management emphasises tasks that resemble the job and competencies a person needs on entry. That supports using relevant work rather than testing an obscure technique the successful candidate could learn after joining. [Source 1]
Agree the expected evidence before choosing a format. Include correctness, reasoning, verification and communication where they matter. Do not award points for building features the brief never requested.
Compare the formats honestly
Each format creates a different view of the candidate’s work. The table below is a planning aid, not a ranking of which method predicts performance best.
| Format | Useful evidence | Main design risk |
|---|---|---|
| Short take-home | Independent decisions, organisation and written explanation | Unlimited polishing or unequal free time |
| Live coding or debugging | Investigation, questions and response to new information | Interview pressure dominating the task |
| Paid trial | Collaboration and decisions over a longer period | Excessive commitment or unclear employment arrangements |
| Existing-work discussion | Ownership, trade-offs and learning from real work | Confidentiality limits or unclear individual contribution |
Choose a primary format and use a short discussion to resolve specific gaps. Requiring a take-home, a live coding round and a trial by default can repeat the same evidence while consuming much more candidate time.
Make a take-home genuinely bounded
A useful take-home has a small task, a stated time budget and a clear stopping point. Corellia’s recommendation is to design an initial exercise around roughly one to two hours of work, then pilot it with someone unfamiliar with the brief. That is a design recommendation, not a legal threshold or a market benchmark.
Supply a working starting point where setup is not part of the assessment. State what to submit, what to leave unfinished and how to explain remaining work. Make clear that you will review the decisions within the stated time, rather than reward a weekend of additional development.
For example, a fictional exercise might ask someone to repair an unreliable endpoint and add a test for the failure. The follow-up can explore how they reproduced the problem, what they deliberately left alone and what they would monitor after release.
Give candidates a reasonable window in which to choose their working time. A submission deadline and the amount of effort required are different things. Ask whether the format is practical before sending it.
Make live work collaborative
Live coding works best when the candidate understands the environment, the task and the interviewer’s role. Provide setup instructions beforehand and let them use familiar tools where the job allows it.
Begin with context and give them time to read. Let questions improve the task rather than treating every clarification as a weakness. If the exercise contains an ambiguity, decide in advance what information the interviewer can provide.
A debugging session can reveal more than building a fresh feature under a stopwatch. Ask the candidate to examine a small failure, form a hypothesis and choose the next observation. You can then change one constraint and discuss the effect.
Record what they did and why it matters. “Checked the failing boundary before changing the implementation” is usable evidence. “Seemed confident” is much harder to compare. Give assessors a consistent rubric and a route for recording uncertainty.
Use a paid trial only when it adds something
A trial is appropriate when a longer shared working period answers a real unresolved question and both sides can make the time. It should have a defined scope, payment, duration, supervisor and decision date.
Do not assume a candidate can take leave from their current job or work evenings. Offer an alternative where feasible, and consider confidentiality or conflict restrictions before inviting them into live work.
Agree ownership of the output, access permissions and how the trial ends. Keep production access proportionate to the task. A limited test environment often gives the evidence needed without exposing customer data or making the candidate responsible for a live incident.
UK government guidance says an unpaid recruitment trial can still create worker status and entitlement to minimum wage, depending on the circumstances. Calling something a test does not settle that question. Longer or productive work needs its employment and payment arrangements resolved before it begins. [Source 2]
Set the AI policy before the task
Candidates should know whether AI tools are permitted, required or excluded, and why. Match that policy to the question being tested.
If the job involves AI-assisted development, allow appropriate tools and assess verification: how the candidate checks generated code, notices incorrect assumptions and explains a change. Ask for a concise account of what they used, rather than intrusive monitoring of their entire working session.
If you need to see an unaided skill, explain that purpose and keep the segment narrow. Do not ban tools in a task designed to resemble the working environment, then praise their use after hiring.
A follow-up discussion is useful in either case. Ask the candidate to modify a requirement, inspect a failed test or explain an important boundary. This reveals understanding more directly than guessing authorship from the style of a submission.
For AI-specific roles, first distinguish AI engineering from research engineering, then select the work sample accordingly.
Reduce avoidable bias and drop-out
A fair process gives comparable candidates comparable opportunities to show the required skill. Share the task purpose, timing, evaluation criteria and adjustment contact in advance. Keep the criteria connected to work rather than personal similarity.
Different candidates may need different arrangements to demonstrate the same competency. Discuss adjustments without making the candidate explain private circumstances to every interviewer. Record any agreed change separately from the assessment of their engineering work.
Measure where people withdraw and ask for a short reason when they are willing to share it. Separate scheduling, task length, compensation and loss of interest in the role. A lower completion rate alone does not tell you which part needs changing.
Review evidence after the first few uses of a new exercise. If capable candidates repeatedly misunderstand the same instruction, improve the instruction. If every assessor reaches a different conclusion, revisit the rubric before testing more people.
Close the assessment with a decision
Assign one person to gather the evidence and decide the next step. Ask assessors to write their observations before discussing the candidate as a group, so the first confident opinion does not set the whole conversation.
Resolve meaningful disagreements with a targeted question, not an automatic extra round. If the evidence already answers the hiring criteria, move to the decision. If it does not, explain exactly what is missing and choose the smallest fair way to investigate it.
The package and role should already be aligned by this stage. Use the salary guide for the compensation conversation and the technical leadership mandate guide where senior ownership is central. A good assessment lets both sides understand the work and make the next decision with better evidence.
If you are planning a technical hire, Start a conversation.
Legal and tax points are summaries; take professional advice.
Sources
- Work samples and simulations, US Office of Personnel Management, checked 31 October 2026.
- Minimum wage eligibility and unpaid work trials, GOV.UK, checked 31 October 2026.
The format comparison, time budget and exercise examples are Corellia Talent’s editorial recommendations. The exercise is fictional.