Skip to content

PROMPTS.md

Latest commit

 

History

History
71 lines (50 loc) · 3 KB

PROMPTS.md

File metadata and controls

71 lines (50 loc) · 3 KB

Prompt Engineering Guidelines

Warning

CRITICAL: DO NOT MODIFY THE OUTPUT STRUCTURE The DR-Engine relies on strict Regular Expression (Regex) patterns to parse the AI's response. Modifying the structural instructions or the format example in the prompt templates WILL BREAK the generating workflow.

Overview

The generic "Question by Question" workflow operates in two distinct phases:

  1. Framework Generation: Creates a structured list of themes and questions.
  2. Answer Generation: Generates detailed answers for each section.

The application parses the output of Phase 1 to construct the database entries for Phase 2.

1. Framework Generation Prompt

File: src/lib/api/prompt-templates.ts (QUESTION_PROMPT_TEMPLATE)

How the Parser Works

The function parseQuestionsResponse in src/lib/api/perplexity.ts looks for specific patterns:

  1. Themes: Must match patterns like I. Title, **I. Title**, or Theme 1: Title.
    • Regex: /^(?:#{1,3}\s*)?(?:\*\*)?\b([IVXL]+|[0-9]+)\.\s*(.+?)(?:\*\*)?$/i
  2. Questions: Must start with a number, bullet, or dash, and must include a question mark.
    • Regex: /^(?:\d+\.|\*|-)\s+(.+)$/ + .includes("?")

✅ Safe to Customize

You may freely modify these sections to change the content and quality of the research:

  • # CORE OBJECTIVES: Change the goal of the research.
  • # ANALYSIS PROTOCOL: Change the methodology (e.g., from "Academic" to "Market Research").
  • Terms & Definitions: You can add domain-specific glossary instructions.

⛔ Do Not Touch

DANGER ZONE: Do not modify or remove these sections:

  • # OUTPUT TYPE: Research questions only.
  • # OUTPUT STRUCTURE: The definitions of "Thematic Matrix" and "Question Hierarchies".
  • # FORMAT EXAMPLE: This provides the few-shot example the AI follows. Changing this will break the parser.

2. Answer Generation Prompt

File: src/lib/api/prompt-templates.ts (ANSWER_PROMPT_TEMPLATE)

This prompt is less brittle structurally but important for quality.

Variables

The engine injects the following variables at runtime:

  • {topic}: The main research topic.
  • {theme_title}: The specific section being processed.
  • {questions}: The list of questions belonging to that section.

✅ Safe to Customize

  • Tone & Style: Change "Academic yet accessible" to "Executive Summary style" or "Blog post style".
  • Length: Adjust "200-400 words" to your preference.
  • Sources: Modify citation rules (e.g., specific domains to include/exclude).

⛔ Do Not Touch

  • Variable Placeholders: Do not remove {questions}, or the AI won't know what to answer.

Troubleshooting

If your workflows are stuck in "Generating Questions" or return empty sections:

  1. Check if you modified QUESTION_PROMPT_TEMPLATE.
  2. Did you change the specific numbering style (e.g., 1. vs One:)?
  3. Did you remove the FORMAT EXAMPLE section?
  4. Revert to the original template and apply changes incrementally.