A model generates its answer one chunk at a time, and each chunk it writes becomes part of what it reads for the next chunk. When a problem takes several steps to solve, the words the model has already written are the only place it can "hold" the intermediate work. So if you ask only for the final answer, it has to jump straight to a result with nowhere to keep the middle steps, and on anything with more than one step it often guesses wrong.
Give it room to write the steps out and the accuracy changes. Each step it writes down feeds the next one, so the model builds the answer on its own working instead of leaping to a conclusion.
Take a small word problem:
A deploy runs every 15 minutes and each run takes 4 minutes. If a run starts
at 09:00, what time does the third run of the hour finish?
Ask for just the answer and a model may blurt out a time that sounds reasonable but is off, because it never worked out when each run starts.
Now ask the same thing but add one line:
Work through it step by step, then give the finish time.
This time it writes something like "first run 09:00, second 09:15, third 09:30, third run finishes at 09:34" and lands on the right answer. Nothing about the model changed. You just let it show the arithmetic instead of skipping to the end.
Writing out the steps helps whenever the answer depends on a chain of smaller results:
For these, the steps are not just for you to read. They are what makes the final answer more likely to be correct.
Next: the exact phrases that ask for this, and when it is not worth the extra words.
Your lab setup
This VM comes preconfigured with your AI keys - just type claude to get started. Your key and URL are already set as $OPENAI_BASE_URL and $OPENAI_API_KEY.