What’s the difference between Chain-of-Thought (CoT) and Chain-of-Decision (CoD) prompting?

What you need to know

Be honest about the term

Chain-of-thought is a well-known, published technique. "Chain-of-decision" is not a single standard method; different teams use it for slightly different things. And Chain of Draft, a 2025 technique, is also abbreviated CoD: it asks the model to reason in very short notes instead of full sentences, to cut tokens and latency. In an interview, ask which one they mean. That alone shows you know the field.

The underlying difference

A decision record looks like this:

{"step": 2, "question": "Has the waiting period for this condition passed?", "options": ["yes", "no", "needs_adjuster"], "choice": "no", "evidence": "policy start 2025-04-01; admission 2025-05-10; waiting period for this condition is 24 months"}

Why teams want the structured form

• Code can validate each decision: is choice one of options? Does evidence cite a real document?
• It maps onto routing and approval rules: needs_adjuster goes to a human.
• Evaluation is per decision, which is easier than scoring a paragraph.
• It produces an audit record for regulated work.

The compromise

Let the model reason freely (or use a reasoning model's hidden thinking), then require a structured decision object as the output. The system consumes the structured object; the prose is for debugging only.

A real-life example

A health insurer's claims agent produced free-form reasoning ending in "APPROVE" or "DENY". Auditors could not check it, and adjusters had to read long paragraphs.

The team changed the output to a fixed list of five decision points: policy active, waiting period, exclusion check, documents complete, amount within limits. Each has options, a choice and evidence that must cite a document ID.

Results: a validator now rejects any record whose evidence cites a document that is not in the claim file; this caught 4% of drafts that had invented a discharge-summary detail. Adjusters review five short fields instead of a paragraph, and review time per claim fell from 6 minutes to about 2.

Follow-up questions to expect

• "When is free-form CoT better?" — For open problems where the useful move is not in any predefined list, such as debugging an unfamiliar incident.
• "How do you enforce the structure?" — Use the provider's structured-output or tool-calling mode with a JSON Schema, then validate in code.
• "Does structure reduce accuracy?" — It can if forced too early, which is why reasoning first and structuring the output second is the common pattern.