Common pitfalls

This page is designed to grow. Every time a student hits something confusing, it becomes an entry. If you're stuck, scan this page first — your problem is probably here.

Starter entries

• Plausible-but-wrong numbers. The agent confidently reports a metric that turns out to be calculated against the wrong column. Always spot-check at least one number against the raw data.
• Hallucinated functions. The agent writes code calling some_library.do_thing() that doesn't exist. Sign: the code crashes on first run with "no such function." Fix: tell the agent which library you actually have, and to verify the function exists before using it.
• Silent failures. The script runs, exits with no error, and produces nothing. Sign: empty output file or a success message with no detail. Fix: ask the agent to add a final "verify" step that prints what was produced.
• Over-prompting on simple tasks. You wrote a page-long prompt for a one-line task. Counterproductive — match prompt length to task complexity.
• Treating prompts as one-shot. First output is 80% right and you start over instead of asking for the 20% fix. Iterate, don't restart.