Chain-of-thought (CoT) prompting asks a language model to show intermediate reasoning before the final answer. Instead of jumping to a conclusion, the model outlines steps—improving accuracy on math, logic, policy interpretation, and multi-constraint planning.

Use Generate a prompt to draft a CoT template for your task, then adapt the steps below.

When CoT helps

CoT is worth the extra tokens when:

The task has **multiple constraints** (budget + timeline + compliance)
You need **auditable reasoning** (support, legal triage, internal ops)
**Arithmetic or conditional logic** appears (discounts, eligibility rules)
The model previously gave confident but wrong one-shot answers

Skip CoT for simple rewrites, short summaries, or creative brainstorming where rigid steps hurt flow.

Basic CoT template

``` Task: [GOAL] Context: [FACTS THE MODEL MUST USE]

Instructions: 1. List the sub-questions you must answer. 2. For each sub-question, show your reasoning in 1–3 sentences. 3. State assumptions explicitly. 4. Give the final answer in [FORMAT]. 5. If information is missing, say what is missing—do not invent data. ```

Example: refund eligibility

**Without CoT:** "Is this customer eligible for a refund?" → often yes/no with invented policy details.

**With CoT:** Model lists purchase date, plan type, refund window, exception flags, then concludes with cited policy clause placeholders.

Variants that work in production

Zero-shot CoT trigger

Append: **"Let's think step by step."** Surprisingly effective on medium tasks; pair with output format constraints.

Structured CoT (JSON)

Require `steps[]` then `conclusion` fields so your app can parse reasoning separately from the user-facing answer.

Self-check pass

After CoT, add: **"Review your steps for contradictions; revise if needed."** Catches inconsistent eligibility logic.

Common mistakes

**CoT on trivial tasks** — adds latency without quality gain
**No format cap** — model writes essays; specify max steps
**Missing grounding** — CoT amplifies hallucinations if context is empty; paste source text
**Skipping verification** — always spot-check high-stakes conclusions

Combine with other techniques

**Role assignment:** Role Assignment Technique
**Few-shot:** show one worked example with visible steps — Zero-Shot vs Few-Shot
**Fundamentals:** What is Prompt Engineering?

Summary

Chain-of-thought prompting trades brevity for reliability on complex reasoning. Ask for explicit steps, cap verbosity, ground in facts, and verify outputs. Start with one high-stakes workflow this week—not every prompt needs CoT.

Next: Prompt Engineering Best Practices for 2026 or generate a CoT prompt.