How to run an FDE mock interview that reveals what to practice next

Run an FDE mock around one task, introduce a requirement the candidate has not rehearsed, and record what they actually do. Afterward, repair the specific gap and try a fresh related task. This produces a more useful study decision than a broad score for confidence or polish. Keep the mock's tool rules and supplied facts explicit, and avoid treating one session as a prediction of an employer's decision.

The format below is an original practice method. Employers use different rounds and policies; confirm yours through the recruiter and the relevant company guide. Use the interview process overview to choose which skill you need to practice first.

Choose one task with an inspectable result

A mock can test coding, decomposition, system design, customer communication or a project deep dive. Trying all of them in 45 minutes usually leaves too little time to investigate any decision.

Start with a task whose expected behavior can be checked. A ticket importer can preserve logical identities under replay. A discovery exercise can distinguish an observed delay from an assumed cause. A customer simulation can end with a bounded recommendation and the correct decision owner.

For an implementation session, use the candidate's own small project or the Foundations capstone. For customer judgment, try the pilot approval conflict. Give the reviewer the relevant facts and acceptance conditions before starting.

Use a schedule that leaves room for the follow-up

The following 45-minute format suits a project defense. Adapt it for a different round rather than pretending it is a standard employer loop.

Agree whether documentation and AI tools are allowed. If they are, observe how the candidate verifies their output. If they are not, enforce that condition consistently. Save feedback for the end unless the session is explicitly coached or intervention is needed for a safety or permission issue.

Change a dependency, not just a number

A useful follow-up exposes an assumption. If the initial importer has globally unique record IDs, introduce two tenants with the same ID. If a result table has complete outcomes, remove two and ask what can still be concluded. If a read-only assistant becomes a message sender, ask which approvals and failure modes change.

The change must be answerable from supplied facts or a reasonable clarification. Do not invent an undocumented constraint halfway through and penalize the candidate for not anticipating it. The reviewer should distinguish finding missing information from failing to use information already available.

For system design, the manual-revision scenario is useful because higher semantic similarity does not resolve equipment applicability. A candidate must revise the retrieval contract. Drawing a larger version of the original architecture would not address that failure.

Record observations that can guide practice

Replace “weak systems thinking” with the action that was missing. Did the candidate define a stable identity? Preserve failed requests in a denominator? Name the team that can approve a data flow? Explain how an operator knows a change worked?

A rubric can organize these notes, but a total score can hide a serious failure. A polished explanation does not compensate for exposing another tenant's records. Keep critical unresolved issues visible and explain why they matter to the task.

If two reviewers disagree, compare the observation and the criterion. One may reward a fast implementation while another expects a clarification the prompt never made necessary. Improve the rubric before assigning meaning to the disagreement.

Use AI as a practice aid with checked feedback

An AI assistant can propose alternate inputs, ask follow-ups and challenge a vague explanation. Give it only material you are allowed to share. Synthetic fixtures are often enough; private customer logs and employer take-homes may have restrictions.

Verify its expected answers. It may suggest an incorrect retry guarantee, miss a denominator problem or reward an answer for repeating the rubric's vocabulary. Its numeric rating has no established relationship to an employer's hiring decision unless that relationship has actually been measured.

Keep the record honest. If the assistant explains the defect before you fix it, the repair is coached. That is useful learning. The next attempt should test the same mechanism with a fresh instance under the tool policy you want to practice. The AI-assisted ownership question shows how to explain assistance and demonstrate understanding together.

Turn the review into the next exercise

Choose one observed gap and write a concrete follow-up. For a missing identity boundary, build a fixture with two tenants and a conflicting replay. For weak customer communication, deliver a shorter update that names the business consequence and decision needed. For unsupported savings, recompute the pilot evidence lab before presenting the recommendation again.

Keep both attempts. Record what changed, which checks ran and where help was required. The portfolio defense lab provides a claim ledger for doing this across several sessions. Use that ledger to choose your next task from the gaps still present in the latest attempt.