Inside the Agent Persona Builder

Most “AI interviewer” tools give you a text box and a submit button. You paste a job description, it generates questions, and whatever the candidate experiences is out of your hands.

We took a different position: the interview is a product surface, and the people who own the role should control it. The Agent Persona Builder is the result — a five-step flow that turns a job description into a working interview agent, with every consequential decision made explicitly rather than inferred.

Watch the demo

Five steps, in dependency order

The builder is ordered by what each step needs from the one before it, not by what’s convenient to ask first.

A progress ring tracks completed sections, and a section only counts as complete when its required fields genuinely validate — not when you’ve merely visited it.

Job details: the input everything else derives from

Four fields, each with a floor that exists for a reason:

• Title — at least 5 characters
• Description — at least 100 characters, up to 1,000
• Experience level
• Top skills — at least 3, at most 5

The skills cap is the one people push back on. Five is deliberate: it forces a decision about what actually matters. Anything broader belongs in the description, which the agent also reads. A list of fifteen “required” skills isn’t a specification, it’s a wish, and it produces an interview that skims everything and establishes nothing.

The 100-character description floor exists because everything downstream depends on it. Generated questions, evaluation criteria and the technical score are all derived from the description. A two-line JD produces a two-line-quality interview.

Evaluation criteria: the part that does the work

This is the step that separates a structured interview from a chat.

Once job details are complete, the builder generates a set of criteria tailored to the role — each a specific, checkable statement about the candidate, with an importance of high, medium or low. You can edit any of them, delete the ones that miss, and add your own.

The step is gated until job details are filled in. That’s intentional: criteria generated from an empty description would be generic, and generic criteria are worse than none because they look rigorous while measuring nothing.

Criteria written to be checkable produce clean results:

Good: “Has led a migration from a monolithic architecture to services”

Weak: “Is a self-starter”

The first can be confirmed or contradicted by something a candidate says. The second cannot, and it will come back as Not covered on every report — which is a signal about the criterion, not the candidate.

Each criterion later appears on the candidate report as Met, Partially met, Not met, or Not covered, so the connection between what you asked for and what you got back stays visible.

Agent persona: what the candidate actually experiences

Everything up to here shapes evaluation. This step shapes the conversation.

Language

Six options, and the two Arabic entries are separate for a reason:

• Arabic — Egypt
• Arabic — Saudi Arabia
• English
• French
• German
• Spanish

Egyptian and Gulf Arabic are different enough that using the wrong one is immediately obvious to a native speaker, and a candidate who notices the mismatch in the first thirty seconds spends the rest of the interview slightly off balance. Treating them as one locale is a shortcut that costs you signal.

Voice

Each language offers several voices, described by style and tone — warm and professional, calm and authoritative, bright and modern, assured and thoughtful. Every one has a preview you can play before committing, from the list or from the detail panel.

Voice is not decoration. It sets how formal the interview feels, and that changes how candidates respond. Our rough guidance: measured, executive-leaning voices for senior and client-facing roles; warmer, brighter ones for high-volume early-career hiring, where drop-off is the bigger risk.

Changing language re-scopes the voice list and keeps your selection if that voice exists in the new language — so switching from English to Arabic mid-setup doesn’t silently reset your persona.

Instructions and custom questions

Two optional levers:

Instructions prompt (up to 100 characters) — a steer on emphasis, for example “focus on problem-solving and cultural fit.” The limit is deliberate. It’s a nudge, not a second job description, and long free-text instructions tend to fight the structured criteria rather than complement them.

Custom questions — up to 8, each between 10 and 500 characters, asked alongside the generated ones. Use them for the thing you always ask and can’t risk being skipped.

One trap worth naming: eight custom questions inside a 10-minute interview leaves almost no room for anything else. Match question count to duration. And if a custom question maps to something you want scored, add a matching evaluation criterion — questions gather evidence, criteria are what appear on the report.

Interview setup: the constraint everything runs inside

Duration is the highest-leverage setting in the builder, and the one most often set without thinking.

Three options, each consuming quota differently — 10, 20 or 30 minutes. Twenty is the sensible default for most screening. Ten works for genuine high-volume first passes with three or four criteria. Thirty earns its cost on senior and deeply technical roles.

The failure mode is always the same shape: eight criteria and eight custom questions inside a 10-minute interview. The agent cannot cover it, and the report comes back with Not covered rows that look like candidate weaknesses and aren’t.

Also here: whether the camera is required, whether the interview is recorded, and how long invitations stay valid before expiring.

ATS integration

The last step connects the agent to where your team already works. Lever is live — pick a posting, enable sync, and candidates and results flow through. More providers are in progress and appear in the picker as they land.

The principle underneath

Every one of these steps is a place where a tool could have guessed and didn’t.

We could infer duration from seniority. We could pick a voice by region. We could generate criteria and never show them. Each of those would shorten setup by a few seconds and remove the hiring team’s ability to explain, later, why a particular candidate was scored the way they were.

An interview that affects someone’s career should be one you can account for. That means the decisions are yours, made visibly, before anyone is interviewed.

---

Ready to build one? Start free, or read the setup walkthrough in Create your first position.