Chain-of-Thought

What It Is

Asks the model to reason step by step before answering. Introduced in Wei et al., 2022, "Chain-of-Thought Prompting Elicits Reasoning in Large Language Models" (arXiv:2201.11903). Self-consistency (Wang et al., 2022, arXiv:2203.11171) samples several chains and takes the majority answer.

When to Use It

Use for maths, debugging, planning and trade-off analysis. On models with built-in extended thinking, ask for reasoning in a separate tag rather than forcing long visible chains.

Blank Template

{{problem}}
Think through this step by step inside <thinking> tags.
Then give the final answer inside <answer> tags.

Five Sample Prompts for Gen AI Professionals

COT-01 - VRAM budget calculation

A {{params}}B model at {{quant}} quantisation with a {{ctx}}-token context on a GPU with {{vram}} GB VRAM.
Think step by step in <thinking>: weight memory, KV cache size, runtime overhead.
In <answer>: does it fit, and what is the largest context that fits?

COT-02 - Root-cause a latency spike

<metrics>{{p50/p95/p99 before and after}}</metrics><changes>{{deploy log}}</changes>
Reason step by step in <thinking> through each possible cause, ruling them in or out with evidence from the metrics.
In <answer>: the most likely root cause and the one test that confirms it.

COT-03 - Token cost forecast

We expect {{users}} users, {{requests_per_user}} requests a day, {{input_tokens}} input and {{output_tokens}} output tokens per request, at {{price_in}} and {{price_out}} per million tokens.
Work it through step by step in <thinking>, including prompt-caching savings at {{cache_hit_rate}}.
In <answer>: monthly cost table for 3 growth scenarios.

COT-04 - Rust borrow-checker error

<code>{{code}}</code><error>{{compiler_error}}</error>
In <thinking>, trace the lifetime of each borrow step by step.
In <answer>: why the compiler rejects it and the smallest idiomatic fix.

COT-05 - Use-case prioritisation

<use_cases>{{list with value, effort, risk notes}}</use_cases>
In <thinking>, score each use case on value, feasibility, data readiness and risk, showing your reasoning.
In <answer>: a ranked list with a one-line reason each and the top pick for a 6-week pilot.

Links

• Up: Frameworks
• Curated by: curation prompts