prompt-engineering

name	prompt-engineering
description	Procedural knowledge for designing, evaluating, and versioning robust LLM prompts.

Skill: Prompt Engineering

Procedural knowledge for designing, evaluating, and versioning robust LLM prompts.

---

Purpose

Enable the AI Expert and Developer roles to construct reliable, secure, and highly-performant prompts that consistently produce the desired output format and behavior from Large Language Models.

---

Prerequisites

• Target LLM (e.g., Claude 3 Opus, GPT-4) and its specific idiosyncrasies known
• Exact expected output format (JSON, XML, Markdown) defined
• Test cases or a golden dataset available for evaluation

---

Procedures

1. Prompt Construction

Step 1: Define the Persona & Context
- Give the LLM a specific role (e.g., "You are an expert PostgreSQL DBA").
- Provide all necessary background context.

Step 2: Clear Instructions & Constraints
- Be explicit about what the model MUST do and MUST NOT do.
- Use numbered lists for complex, multi-step reasoning.

Step 3: Output Formatting
- Provide exact schema requirements.
- Use XML tags (e.g., `<output></output>`) to structure complex inputs and expected outputs.
- Ask for reasoning *before* the final answer (Chain of Thought) to improve accuracy.

Step 4: Few-Shot Examples
- Provide 2-3 high-quality examples of inputs and their ideal outputs.

2. Mitigation Strategies

Step 1: Hallucination Prevention
- Instruct the model to say "I don't know" or "Data not present" if the answer isn't in the context.
- Require citations or direct quotes from the source text.

Step 2: Prompt Injection Defense
- Delimit user input strictly using XML tags (e.g., `<user_input>`).
- Instruct the model to ignore any instructions hidden within the user input block.

---

Common Pitfalls

1. Vague Instructions: Leaving ambiguity leads to inconsistent outputs.
2. Missing Constraints: Failing to tell the model what not to do.
3. Premature Output: Forcing the model to output the final answer before reasoning through it.
4. No Version Control: Treating prompts as disposable text instead of code.

---

Invocation

"Apply the prompt-engineering skill from .ace/skills/prompt-engineering/SKILL.md
to design the extraction prompt for the invoice processing feature."