Applied AI / Forward Deployed Engineer · Prompt Engineering · Lesson 3 of 4
Chain-of-thought & reasoning control
Chain-of-thought (CoT) prompting asks the model to reason step by step before answering. For multi-step problems (math, logic, complex extraction), it substantially improves accuracy because the model uses generated tokens as a scratchpad.
- Simple trigger: "Think step by step, then give the final answer."
- For clean output, have it reason in a hidden/section then emit a final structured answer — or use a model with built-in reasoning.
- Reasoning models (built-in 'thinking', now standard on frontier APIs and usually dialed with a reasoning-effort setting) do this internally — telling one to 'think step by step' is obsolete and just adds cost. Save manual CoT for non-reasoning models.
Don't parse the reasoningIf you show your reasoning and your final answer together, extract the final answer explicitly (e.g. after a marker or in a JSON field). Don't regex the whole blob. And more reasoning = more tokens = more cost/latency — use it where it pays off.
Nuance that impresses"For arithmetic-heavy or multi-hop tasks I'd enable step-by-step reasoning, but for a reasoning model I'd rely on its internal thinking and just constrain the final output — no need to pay for verbose CoT I then have to parse."
◆ Lock it in
- CoT ("think step by step") boosts multi-step accuracy by giving the model a scratchpad.
- Separate reasoning from the final structured answer; extract the answer explicitly.
- Reasoning models think internally — don't force manual CoT; mind the token cost.
Feynman drill — say it out loudWhy does asking a model to 'think step by step' improve accuracy on a math word problem?
Step 1 rate your confidence · Step 2 pick your answer
For a multi-step logic task, chain-of-thought helps mainly because:
Step 1 — how sure are you?