Talent Systems — Science Team
Interviewer Prompt

Full Prompt Anatomy

Every piece of data that goes into the interviewer's system prompt.

The interviewer prompt is assembled dynamically at interview start. It combines rubric configuration (science-owned) with per-posting and per-candidate data (dynamic).

Structured (Spine) Mode

Built by lib/interviewer-prompt.ts, delivered to Retell GPT-4.1 via retell_llm_dynamic_variables.

Science-Owned Fields (editable via admin UI)

FieldConfig KeyWhat It Controls
System identitysystem_identityHow the AI identifies itself. Tokens: {companyName}, {role}
TonetoneConversational style
Probing styleprobe_styleHow to ask follow-ups
PacingpacingWhen to let candidate finish, when to redirect
Opening scriptopeningHow to start the interview
Closing scriptclosingHow to end the interview
Question spine instructionsquestion_spine_instructionsRules for following the spine
Probing triggersprobing_triggersWhen to ask follow-ups
Max probes per questionmax_probes_per_questionHard cap on follow-ups (number)
Probe typesprobe_typesAllowed follow-up categories
Escalation rulesescalation_rulesWhen to probe harder vs. move on
Depth configdepth_configDuration targets per depth level (numbers)
Max resume charsmax_resume_charsHow much resume text the AI sees (number or null for full)
Hard ruleshard_rulesAbsolute constraints (add/remove list)

Dynamic Fields (per posting/candidate)

FieldSourceEditable?
JD textpostings.jd_summaryBy employer (posting form)
Question spinepostings.question_spineBy employer (posting form)
Resume textcandidates.resume_text (truncated)By candidate (upload)
Resume gapscandidates.resume_gapsAuto-detected by AI
Custom notespostings.custom_notesBy employer (hiring manager context)

Auto-Appended Rules

Two rules are automatically appended based on posting config:

  • "Keep the interview to approximately N questions plus probes" (from spine length)
  • "You may ask up to N follow-up probes per question" (from max_probes_per_question)

Adaptive Mode

Same structure as above, plus additional fields. See Adaptive Mode for the full field list.

Key difference: receives full competency anchors (1-5) and competency weights instead of a question spine.

On this page