Prompt Engineering for Developers: 7 Patterns That Actually Work

Most “prompt engineering” advice is vague nonsense. “Be specific.” “Provide context.” Thanks, very helpful.

Here are 7 concrete patterns I use daily that make a measurable difference when working with AI coding tools.

Pattern 1: The Constraint Sandwich

Structure: Task → Constraints → Task reminder

Why it works: AI models tend to forget constraints mentioned at the beginning by the time they start generating. Repeating the core constraint at the end keeps it in focus.

Without this pattern: You’ll get a function that uses email-validator package, returns just a boolean, and is 60 lines long.

Pattern 2: The Negative Example

Structure: Show what you DON’T want, then what you DO want.

Why it works: AI models learn from millions of code examples, including bad ones. Showing the anti-pattern explicitly prevents the model from falling into the most common trap.

Pattern 3: The Role Lock

Structure: Assign a specific expert role with domain knowledge.

Why it works: The role constrains the model’s response distribution. A “senior Python developer” will suggest different things than a generic assistant — more idiomatic code, better patterns, awareness of production concerns.

Pro tip: Match the role to your actual needs. Don’t say “world’s best developer” — say “backend developer experienced with Django and PostgreSQL” for more relevant suggestions.

Pattern 4: The Chain of Files

Structure: Feed files in dependency order with clear labels.

Why it works: AI needs context to understand code. But dumping your entire codebase doesn’t help — it dilutes the signal. Feeding files in dependency order with clear labels gives the model a mental map of your system.

The label “BUG IS HERE” focuses attention on the right file while still providing necessary context from dependencies.

Pattern 5: The Output Template

Structure: Show the exact output format you want.

Include tests for: happy path, zero values, negative values, boundary conditions, and invalid input types.

I need a REST API for a todo app. Let’s build it step by step.

Step 1: Generate just the data models (SQLAlchemy).

• Todo: id, title, description, status, created_at, updated_at
• User: id, email, hashed_password, created_at Wait for my approval before proceeding.

Good. Step 2: Generate the CRUD operations for Todo model.

• Use async SQLAlchemy
• Include pagination for list endpoint
• Include filtering by status Wait for approval.

Bug: API returns 500 error intermittently (about 1 in 20 requests).

Stack trace: [paste stack trace]

Context:

• This started after deploying commit abc123 (added connection pooling)
• Only happens under load (>50 req/sec)
• Database connections look normal in monitoring
• No memory leaks detected

Already tried:

• Increasing connection pool size (no change)
• Adding request timeout (no change)
• Rolling back connection pooling (fixes it, but we need pooling for performance)

What are the most likely causes? Rank by probability.