In short
Asking a model to show its steps before its answer. Better results on hard problems, and a visible place to find where it went wrong.
Chain of thought is the practice of having a model reason in the open: list the facts, work through the logic, then conclude. On problems with several steps it produces better answers than jumping to a conclusion, and it leaves a trail.
The trail is the useful part for a business. When an agent's decision is wrong, the reasoning shows which step went wrong, and that step points at the sentence in the instruction file that needs to change. Output is information, one level deeper.
Current models do much of this on their own. The habit still matters in how instructions are written: ask for the reasoning when the decision is consequential.
Example
What it looks like.
Instruction: before classifying a lead as qualified, list the three criteria and say which are met. The list is kept with the classification so a person can check it in five seconds.Related
Read next.
- output is informationA bad result is data about what the model understood. Read it, adjust, go again. Beginners stop at two tries. Builders go fifteen.Read the definition
- promptWhat you send a model to get something back. In a chat, a message. In an agent, the instruction file plus the situation. Save them. Improve them. Never retype them.Read the definition
- evaluationGrading the work. Measuring whether an agent's output meets the bar, so changes can be tested instead of hoped for.Read the definition