prompt-engineer

Prompt Engineer Skill

Goal

Transform vague, incomplete, or underperforming prompts into precisely crafted instructions that reliably produce the intended output from an LLM.

Prompt Engineering Fundamentals

The Six Dimensions

Every strong prompt addresses:

1. Role — Who is the model? What expertise does it bring?
2. Task — What exactly should it do? (verb + noun: "analyze", "write", "classify")
3. Context — What background does the model need to succeed?
4. Format — What should the output look like? (length, structure, examples)
5. Constraints — What should it avoid? (tone, topics, length limits)
6. Examples — Few-shot examples for consistency (optional but powerful)

Steps

1. Diagnose the existing prompt (if one exists)

If the user has a prompt that "doesn't work":

• What output is it producing?
• What output is desired?
• Which of the six dimensions is missing or weak?

Name the specific problem before prescribing a fix.

Success criteria: You can say "this prompt fails because [specific reason]."

2. Gather requirements

Ask (or infer from context):

• What model will this run on? (GPT-4, Claude, Gemini, Llama — each has different strengths)
• Is this a system prompt, user prompt, or both?
• Will it be used once or repeatedly at scale?
• What's the most common failure mode the user wants to prevent?

Success criteria: You know the model, use case, and primary failure mode.

3. Write the prompt

Apply these techniques based on the use case:

For complex reasoning tasks:

• Add "Think step by step before answering"
• Use XML tags to separate sections: <context>, <task>, <format>
• Include a "before you respond, check:" section

For consistent formatting:

• Provide an exact output template with placeholders
• Add a negative example: "Do NOT format it like this: [bad example]"

For role-playing/personas:

• Define expertise: "You are a senior [role] with [X] years experience in [domain]"
• Define behavioral constraints: "You communicate in [style]. You never [bad behavior]."