Engineering Craft › AI-Assisted Development

Prompting for Code

Giving an AI enough context and constraints to produce useful code.

Also known as: prompt engineering for code, writing prompts for coding assistants

The quality of what an AI coding assistant produces depends a lot on what you give it. Prompting for code means supplying enough context and constraints that the answer is useful, instead of generic.

A weak prompt:

write a function to process orders

A better one:

In orders/service.py (Python 3.12, SQLAlchemy 2), add cancel_order(order_id, user). It should: raise NotFound if the order doesn’t exist, raise Forbidden if it belongs to another user, refuse to cancel orders already shipped, and set status to cancelled inside a transaction. Follow the style of refund_order in the same file. Add pytest tests next to the existing ones.

What to include

• The goal and the context: what it’s for, what already exists, where the change goes. Paste or point to the relevant code (context windows).
• Constraints: language and versions, libraries to use or avoid, style, performance, security rules.
• Behavior and edge cases: inputs, outputs, errors.
• Examples: sample input and expected output, or an existing function to imitate.
• The format you want: just the diff, a plan first, tests included.

Habits that work

• Work in small steps. Ask for one change, check it, then continue. Big requests produce big mistakes.
• Ask for a plan first on anything non-trivial, and correct it before code is written.
• Show errors fully: paste the whole message and stack trace, and what you expected.
• Say what’s wrong when it misses, specifically, instead of repeating the request.
• Keep project rules in a file the assistant reads each session (agent instructions file).
• Start fresh when a long conversation drifts.
• Verify the result: run it, test it, read it (reviewing AI code).

Cautions

• Don’t paste secrets or private data into tools you aren’t cleared to use.
• Detail doesn’t guarantee correctness. It can still hallucinate.
• If you can’t describe the problem clearly, work on understanding it first. The prompt is a mirror of your thinking.