ChatGPT Prompt Engineering Principles: Chain of Thought Prompting

Chain-of-thought prompting—breaking a riddle into a sequence of sub-problems, solving each in order, then combining the results—can dramatically improve an LLM’s accuracy on tasks that fail when answered “from the last sentence.” The core idea is simple: when a question depends on intermediate facts (location, identity, relationships, hidden constraints), forcing the model to enumerate the needed steps reduces guesswork and helps it avoid jumping to an unsupported conclusion.

The approach is presented as a reusable “principle,” not a universal fix. It’s most effective when the problem naturally decomposes into a chain of reasoning steps. The method starts by instructing the model to list, systematically and in detail, the sub-problems required to reach the final answer. After that checklist is produced, the model is prompted to solve each sub-problem—often using a “highest probability” preference when certainty isn’t available—before wrapping up with the final response.

A first example tests the method on a multi-hop riddle about Michael, a 31-year-old American, who is in France looking at a famous museum painting. The riddle then links the painting’s artist to Michael’s favorite childhood cartoon character, and asks for the country of origin of an object the character holds. A straightforward attempt to answer directly fails repeatedly. With chain-of-thought prompting, the model first identifies the needed intermediate pieces: Michael’s location (museum), the most famous painting, the painting’s artist, the cartoon character, and the object the character holds. It then fills them in step-by-step: Michael is inferred to be at the Louvre in France; the painting is the Mona Lisa by Leonardo da Vinci; the cartoon character is guessed as Teenage Mutant Ninja Turtles’ Leonardo (chosen as the most likely match given the Renaissance-artist naming pattern); and the object is a katana. With those components assembled, the final question—country of origin of the katana—lands on Japan. The key takeaway is that the model can reach a correct answer only after being guided through the riddle’s dependency chain.

A second example contrasts chain-of-thought behavior against a zero-shot response. The riddle: a ball is placed into a small box missing the bottom, carried to a postal office, placed into a bigger box, and shipped to a friend in New York; the question asks where the ball ends up. A zero-shot answer claims the ball is in the bigger box, which the narrator rejects because the small box’s missing bottom should allow the ball to fall out. When chain-of-thought prompting is used, the model enumerates additional uncertainties and intervening actions—whether the ball falls out when the box is moved, where it could land during transit, and how the end state is defined. It ultimately concludes the ball most likely fell out of the small box either in the office or on the way to the postal office, and after forcing a single final choice, it selects the office as the most probable location. The result is treated as an improvement because it aligns with the physical constraint introduced by the missing bottom and shows more careful handling of ambiguous steps.

Overall, the transcript frames chain-of-thought prompting as a practical prompt-engineering technique for riddles and multi-step problems: it turns “one-shot guessing” into structured decomposition, which can convert previously unsolvable questions into solvable ones.