Using AI Coding Assistants Well (and Safely)

What they are good at

• Boilerplate: DAG skeletons, schemas, CLI wiring, test scaffolding
• Explaining unfamiliar code or an error message
• Translating between SQL dialects or from pandas to PySpark
• First drafts of documentation and tests

What to watch

• Plausible but wrong code: APIs that do not exist, wrong defaults, subtle logic bugs
• Outdated library versions
• Over-engineered solutions to simple problems

A workflow that works

1. Give context: the file, the schema, the error, the constraint ("must be idempotent").
2. Ask for small, specific changes rather than whole systems.
3. Run it: tests, linters, a dry run on sample data.
4. Read the diff before committing, as you would review a colleague's code.

Write a PySpark function that keeps the latest row per order_id by updated_at.
Input schema: order_id string, updated_at timestamp, amount decimal(18,2).
Include a pytest with a duplicate key and a null updated_at.

Protect secrets and data

• Never paste credentials, tokens or customer data into prompts.
• Check your organisation's policy on which tools may see source code.
• Review generated code for hard-coded secrets and unsafe SQL string building.