Prompt Engineering for Developers: The Patterns That Actually Work

The difference between useful AI code and useless AI code is usually the prompt. Not the model. Not the tool. The prompt.

Here are the patterns that consistently produce good results.

Pattern 1: Provide the Data Shape First

Before asking for any function, describe the data it operates on. 'The User object looks like this: [paste your TypeScript interface or example JSON].' This one change eliminates 80% of hallucinated field names and wrong assumptions.

Pattern 2: Specify Your Error Handling Requirements

AI defaults to happy-path code. Explicitly say: 'Throw a ValidationError for invalid input. Return null if the record doesn't exist. Log errors with the format we use in this codebase.' Without this, you get code that silently swallows errors or crashes unexpectedly.

Pattern 3: State What You Don't Want

Negative constraints are powerful. 'Don't use any external libraries for this.' 'Don't modify the function signature.' 'Don't add TypeScript generics — keep the types simple.' AI will often reach for complexity when simplicity is right. Explicitly ruling things out saves iteration.

Pattern 4: Ask for the Explanation First

For complex problems, ask the AI to explain its approach before writing code. 'Describe how you'd implement this, then write the code.' This surfaces wrong assumptions early. If the approach is wrong, you redirect before it writes 50 lines of bad code.

Pattern 5: Give It a Working Example to Extend

Paste existing code from your project and say 'add X to this.' AI is far more accurate when extending real code than when generating from scratch. It matches your patterns, uses your variable names, follows your conventions. This is the fastest way to get code that actually fits into your project.

The Hardest Pattern: Knowing When to Stop

Sometimes AI just doesn't know the right answer. If you've iterated three times and the code is still wrong in the same way, it might not know the correct approach. Time to read documentation, search Stack Overflow, or just write it yourself. AI tools sometimes confidently produce wrong solutions repeatedly. Recognize that pattern and stop earlier.