prompt-engineering

Prompt Engineering

Broad authority on eliciting reliable, high-quality outputs from any LLM through careful prompt design. Covers system prompt architecture, reasoning techniques, tool-use orchestration, context management, and model-specific behavioral differences. Apply this skill whenever the output quality of an LLM needs to be improved, structured, or made robust — before reaching for fine-tuning.

Core Concepts

System Prompt Architecture

A well-structured system prompt has four zones: (1) role and persona, (2) task framing and constraints, (3) output format specification, (4) edge case handling. Order matters — most models weight earlier instructions more heavily.

You are a senior financial analyst at a hedge fund. You specialize in distressed debt analysis.

## Task
When given a company's 10-K excerpt, produce a structured credit risk assessment.

## Rules
- Base every claim on text explicitly present in the excerpt; never hallucinate figures.
- If information is absent, write "Not disclosed" rather than inferring.
- Use only the output format below. Do not add sections.

## Output Format
<credit_assessment>
  <rating>Investment Grade | Speculative | Distressed</rating>
  <leverage_ratio>[number]x or Not disclosed</leverage_ratio>
  <key_risks>
    <risk>...</risk>
  </key_risks>
  <recommendation>Hold | Buy | Avoid</recommendation>
</credit_assessment>

XML tags are superior to markdown headers for structured outputs because they survive nested generation without ambiguity. Claude responds especially well to XML. GPT-4o and Gemini 1.5 Pro both handle JSON mode natively and can be instructed to output JSON schema-conformant objects.

Few-Shot Examples

Few-shot examples are the most reliable technique for format compliance and tone calibration. Place examples after task description, before the actual input.

## Examples

Input: "The quarterly revenue declined 12% YoY."
Output:
<sentiment>Negative</sentiment>
<magnitude>Moderate</magnitude>
<entity>Revenue</entity>

Input: "We beat consensus estimates by $0.18 per share."
Output:
<sentiment>Positive</sentiment>
<magnitude>Strong</magnitude>
<entity>EPS</entity>

---
Input: {{user_text}}
Output: