Prompt Engineering

Skill: Prompt Engineering

Trigger

When the user needs to write, improve, or review prompts, system instructions, agent definitions, Cursor rules, or any text intended to guide AI behavior.

Steps

1. Clarify the goal — Identify what the prompt should make the AI do (generate code, review, explain, follow a pattern). 2. Define the persona — Specify the role, expertise level, and constraints (e.g. "You are a senior backend engineer specializing in Node.js"). 3. Structure the prompt — Organize with clear sections:

• Context: Background the AI needs
• Instructions: What to do (use imperative mood)
• Constraints: What NOT to do (use NEVER/ALWAYS for hard rules)
• Output format: Expected shape of the response (code, bullet points, table)
• Examples: One good and one bad example when possible

4. Apply prompt patterns — Use established techniques:

• Few-shot: Provide 2–3 examples of desired input/output
• Chain of thought: Ask the AI to think step-by-step for complex reasoning
• Negative constraints: Explicitly state anti-patterns to avoid
• Structured output: Specify exact format (JSON, markdown, code block)

5. Review and iterate — Test the prompt; refine based on output quality.

Rules

• Use imperative mood for instructions ("Use X", "Never do Y")
• Bold critical constraints: ALWAYS, NEVER, CRITICAL
• Keep instructions atomic — one rule per bullet point
• Order from most important to least important
• Use tables for structured data (comparisons, configurations)
• Include escape hatches: what to do when the instruction doesn't apply
• Avoid vague language ("try to", "maybe", "consider") — be decisive
• Test with edge cases: what happens if the user asks something unexpected?

Prompt Quality Checklist

• [ ] Clear role/persona defined
• [ ] Specific, actionable instructions (not vague guidance)
• [ ] Hard constraints marked with ALWAYS/NEVER
• [ ] Output format specified
• [ ] At least one example (good or bad)
• [ ] Edge cases addressed
• [ ] No conflicting instructions
• [ ] Concise — removes filler without losing meaning

Anti-patterns

• Walls of text without structure — use headers and bullet points
• Conflicting rules (e.g. "be concise" AND "explain everything in detail")
• Too many rules — prioritize; AI follows top rules more reliably
• Vague instructions: "write good code" vs "use TypeScript strict mode, define interfaces for all inputs"
• Missing context: assuming the AI knows your project without telling it

Prerequisites

• [ ] Clear understanding of the desired AI behavior
• [ ] Knowledge of the target audience (developer, reviewer, end user)
• [ ] Examples of good and bad outputs

Completion Checklist

• [ ] Prompt has clear structure (context, instructions, constraints, format)
• [ ] Critical rules use ALWAYS/NEVER
• [ ] Tested with at least one real scenario
• [ ] No conflicting instructions
• [ ] Concise (< 500 words for rules, < 1000 words for agents)

If Step Fails

• AI ignores a rule: Move it higher in the prompt; bold it; add "CRITICAL"
• Output is wrong format: Add an explicit example of the expected format
• AI hallucinates: Add "Only use information from the provided context"
• Too verbose: Add "Keep responses under N sentences/lines"
• Too terse: Add "Explain your reasoning" or "Include code examples"

Example

Task: Create a Cursor rule for error handling. 1. Persona: "You are a backend engineer ensuring consistent error handling." 2. Structure: Section for error classes, section for HTTP mapping, section for logging. 3. Constraints: "NEVER expose stack traces to API consumers", "ALWAYS include correlationId". 4. Format: .mdc file with YAML frontmatter, markdown body. 5. Test: Apply to a sample handler and verify output matches expectations.