prompt-engineer

Prompt Engineer

Transform vague instructions into production-grade AI prompts.

Workflow

1. Understand the intent Ask the user (or infer from context):

   • What is the AI supposed to do? (task)
   • Who will use it? (audience)
   • What model will run it? (OpenAI, Claude, Llama, etc.)
   • What format should the output be? (JSON, markdown, free text)
   • Any constraints? (length, tone, safety)

2. Define the role and context Write a system message that establishes:

   • Who the AI is (role)
   • What it knows (context/expertise)
   • What it should NOT do (constraints)
   • How it should respond (tone, format)

   You are a senior code reviewer at a fintech company. You review pull requests
   for security vulnerabilities, performance issues, and maintainability.
   You are direct and specific — cite exact line numbers. You never approve
   code with SQL injection or XSS vulnerabilities.
   

3. Create the user message template Design a structured input format:

   Review this pull request:
   
   **Title:** {{pr_title}}
   **Description:** {{pr_description}}
   **Diff:**
   

   {{diff}}

   
   Focus on: {{focus_areas}}
   

4. Add few-shot examples Create 2-3 input/output examples that demonstrate:

   • The expected quality and format
   • Edge cases the model should handle
   • The boundary between "in scope" and "out of scope"

5. Define output structure Specify the exact format:

   {
     "verdict": "approve | request_changes | comment",
     "summary": "One-sentence overall assessment",
     "findings": [
       {
         "severity": "critical | warning | suggestion",
         "file": "path/to/file.ts",
         "line": 42,
         "issue": "Description of the issue",
         "fix": "Suggested fix"
       }
     ]
   }
   

6. Add guardrails

   • Token budget guidance ("keep responses under 500 tokens")
   • Hallucination prevention ("only reference code in the provided diff")
   • Safety boundaries ("never generate executable code in reviews")
   • Fallback behavior ("if the diff is too large, summarize by file")