Scorer Prompt Anatomy

The scorer runs after every interview, analyzing the transcript against the competency framework to produce a scorecard.

Built by: lib/scorer-prompt.ts → buildScorerPrompt()
Model: Claude Haiku (dev) / Claude Sonnet (prod)
Admin page: /admin/rubrics

What the Scorer Receives

Field	Config Key	Purpose
System identity	`system_identity`	How the scorer identifies its role. Token: `{role}`
Scoring philosophy	`scoring_philosophy`	Anchoring principles ("Score conservatively...")
Ownership rule	`ownership_rule`	How to handle collective vs. individual language
Specificity rule	`specificity_rule`	How to handle vague answers
Quantification bonus	`quantification_bonus`	How metrics affect borderline scores
Coached response flag	`coached_response_flag`	How to detect rehearsed answers
Adverse impact guidance	`adverse_impact_guidance`	What to flag for review
Scoring rules	`scoring_rules`	Add/remove list of additional rules
Advance threshold	`advance_threshold`	Weighted score ≥ this → "Advance" (default: 3.5)
Hold threshold	`hold_threshold`	Weighted score ≥ this → "Hold", else "Decline" (default: 2.5)

Field	Source	Identity Visible?
Competency names	`postings.required_competencies`	No
Competency weights	`postings.required_competencies`	No
Competency anchors (1-5)	`competency_framework.anchors`	No
Integrity flags	`integrity_flags`	No identity — just flag metadata
Transcript	Retell, name-scrubbed	No — `[CANDIDATE]`

Field	Purpose
Output format (JSON schema)	Defines the scorecard structure — scores, evidence, strengths, concerns, recommendation