Prompt Engineer

Problem

When a prompt still produces missed instructions, format drift, context loss, or inconsistent outputs across runs, the issue is often not the model but the prompt structure: role, constraints, examples, failure boundaries, and output format are not separated. Hallucination also comes from missing “cite only / do not invent” anchors and lacking evidence-based iteration.

How It Works

This skill treats prompt engineering as a reusable design path:

• Use system prompt architecture to define role, task, constraints, and prohibited behavior
• Use chain-of-thought design to force analyze, then decide, then answer
• Use few-shot strategies with high-quality examples to stabilize format and style
• Use evaluation-driven iteration to diagnose bad cases and update prompts

v3.0.0 adds 30-second onboarding, a QP error code system, and real samples, making it easier to tell whether the failure is unclear role, weak examples, conflicting constraints, or missing output validation.

Boundaries

It is useful for refining existing prompts, designing prompts for new tasks, and tracing unstable or hallucinated outputs, especially for JSON, reporting, classification, extraction, and reasoning tasks. If the model lacks required capability, the context is truncated, or the system has no real evaluation samples, prompt work can improve results but cannot fully eliminate failure.