Prompt Writing Guide

# Prompt Writing Guide

This document provides reusable, high-fidelity prompt templates, frameworks, and guidelines on how to author and configure production-grade, deterministic AI prompts. It aligns prompt designs with the [AI Engineering Standards](file:///d:/projects/Nexulyt-AI-OS/standards/ai-engineering-standards.md).

---

## 1. Overview

* **Purpose**: Provide developers with a systematic methodology to write instructions that minimize LLM hallucinations, enforce output formats, optimize system boundaries, and regulate context usage.
* **When to Use**: When creating new prompt configurations, writing agent system profiles, constructing prompt-based templates, or refining LLM interaction loops.
* **Inputs**: Target tasks, framework preferences (e.g., CO-STAR), output schema definitions, variables configurations, and context scopes.
* **Expected Outputs**: Production-grade system prompts, few-shot examples layouts, context boundary structures, and instruction filters.
* **Best Practices**:
  - Leverage structured XML tags (`<identity>`, `<instructions>`, `<constraints>`) to separate prompts sections.
  - Define explicit output models (JSON schemas, YAML) to enable automated downstream parsing.
  - Run prompts through systematic evaluations using representative benchmark sets before committing to production.
* **Common Mistakes**:
  - Utilizing vague descriptors like "be helpful" or "avoid bad quality code" instead of concrete, actionable rules.
  - Hardcoding variables in the prompt instead of using standard templating variables (e.g., `{{VARIABLE_NAME}}`).

---

## 2. Prompt Engineering Frameworks

### CO-STAR Prompting Framework
Use the CO-STAR structure to construct comprehensive prompts:
* **Context (C)**: Provide background on the task or scenario.
* **Objective (O)**: State the exact task the model must perform.
* **Style (S)**: Define the tone and writing persona (e.g., technical analyst).
* **Tone (T)**: Establish the execution vibe (e.g., direct, objective).
* **Audience (A)**: Detail who will consume the output (e.g., backend developers).
* **Response (R)**: Enforce the precise format (e.g., JSON schema, Markdown tables).

### CREATE Prompting Framework
Use the CREATE structure for task-oriented outputs:
* **Character (C)**: Define the role or expert persona.
* **Request (R)**: Clearly state the goal or request.
* **Examples (E)**: Embed few-shot input-output pairs to guide the model.
* **Adjustment (A)**: Outline negative constraints (what *not* to do).
* **Type (T)**: Enforce the output structure (e.g., code diff, checklist).
* **Execution (E)**: Guide the logical path the model should take.

---

## 3. Prompt Templates

### System Prompt & Meta-Prompt Blueprint