Chain-of-thought (CoT) prompting asks a model to state intermediate reasoning steps before giving a conclusion. For multistep decisions, the resulting explanation makes the recommendation easier to inspect, challenge, and correct. It is especially useful when a standard, nonreasoning model must weigh several signals, or when a task requires a domain-specific reasoning pattern. section 9.1
Choose the least elaborate prompt that gives the team enough visibility:
- Use a direct prompt when a short recommendation is sufficient; it may leave the weighting of individual signals unclear. section 9.1.1
- Use zero-shot CoT to request a step-by-step assessment of each signal against its threshold before an overall recommendation. section 9.1.1
- Use few-shot CoT when an example must demonstrate both the domain logic and the desired reasoning structure. section 9.1
Canary rollout decision
A release engineer must decide whether to continue a 20% canary rollout. The checkout-api error rate is 3.2%, against a 0.8% baseline and a rollback threshold of more than 3% for at least five minutes. Database connection-pool utilization is 78%, below its 85% alert threshold; a noncritical downstream service has intermittent timeouts. A structured prompt assigns the release-engineer role, supplies these signals, asks for each signal to be assessed against its threshold, and requests a GO or NO-GO recommendation. The intermediate assessment exposes which signals drive the decision. It also gives an engineer a place to check whether the error rate—not merely the rollout—has remained above threshold for five minutes. section 9.1.1
Do not add explicit CoT instructions merely because a task is difficult. If a reasoning or thinking model is available, enable its reasoning mode at the model or API level instead; additional CoT instructions are described as unnecessary token cost. Review an explicit reasoning trace as an argument to verify, not as a substitute for checking its inputs and thresholds. section 9.1 section 9.1.1
