Prompt Engineering

Table of contents

• Purpose
• Prompt contract
• Structured output
• Evidence grounding
• Provider abstraction
• Localization
• Safety and review
• Versioning and evaluation

Purpose

Prompts translate structured RetainAI context into bounded, reviewable advisor output.

Prompt contract

Prompts should define:

role
task
input schema
allowed evidence
output schema
failure behavior
safety constraints
locale

Structured output

Prefer validated structured responses:

{
  "recommendation_key": "advisor.action.review_manager_context",
  "rationale_keys": [],
  "evidence": [],
  "limitations": [],
  "confidence": null
}

Evidence grounding

Provider output must reference supplied evidence and must not invent employee, resume, psychometric, or organizational facts.

Provider abstraction

Supported adapter states:

disabled
gemini
bedrock

Prompt templates should not depend on a provider-specific response format.

Localization

Prompt and explanation architecture must support:

en
es

The backend should return stable message keys and variables where possible. The dashboard renders localized text.

Safety and review

Outputs remain advisory and reviewable. Prompts must not authorize autonomous high-impact employment decisions.

Versioning and evaluation

Track:

prompt version
provider
model identifier
locale
input schema
output schema
evaluation dataset
quality results

Related guides

• Modeling
• Dashboard
• Architecture
• Documentation index