Prompting vs Modeling: Where to Fix the Problem

TL;DR

• Prompting can’t fix ambiguous data models.
• Fix the model first, then refine prompts.

The problem

• Teams try to patch model issues with prompts.
• AI still returns inconsistent answers.

Why it matters

• Prompts are fragile; model changes are durable.
• Model fixes improve all tools at once.

Symptoms

• Prompt tweaks help one question but break another.
• AI still struggles with definitions.

Root causes

• Ambiguous measures and weak metadata.
• Lack of semantic contracts.

What good looks like

• Model definitions are clear, prompts are simple.
• Prompting is used for formatting, not semantics.

How to fix

• Identify root model issues.
• Improve definitions, metadata, and context stability.
• Use prompts for output structure only.

Pitfalls

• Over‑engineering prompts as a substitute for model fixes.
• Ignoring evaluation results.

Checklist

• Model issues addressed first.
• Prompting used for formatting.
• Evaluation shows improved consistency.

Framework placement

Primary ARF layer: AI Readiness & Interoperability. Diagnostic bridge: data-movement-reliability, semantic-reliability, execution-reliability, change-reliability.