Turn a plain question into a reasoning prompt
Asking a model to “just answer” a multi-step problem often gets you a confident wrong answer. Chain-of-thought prompting fixes this by asking the model to reason before it concludes. This tool takes your plain question and wraps it in three proven CoT patterns — step-by-step, think-aloud and scratchpad — so you can copy whichever fits and let the model work the problem out loud.
How it works
Each style is a tested phrasing template. Step-by-step instructs the model to break the problem into explicit numbered steps and only then state the answer. Think-aloud invites freer narrated reasoning, which can surface considerations a rigid list misses. Scratchpad asks the model to do its working inside a clearly marked area and then give a clean final answer beneath it — useful when you want the reasoning available but the deliverable tidy. Your question is inserted into each template and the finished prompts are ready to copy.
Tips and notes
Reach for CoT on genuinely multi-step work — arithmetic, logic puzzles, planning, debugging, anything where intermediate steps matter. Skip it for simple factual lookups, where it only adds cost and latency. If you need a machine-parseable result, the scratchpad style is best because it separates the reasoning from a clearly delimited final answer you can extract. Note that some newer reasoning-optimized models already reason internally, so heavy explicit CoT instructions can be unnecessary or even counterproductive there — test both. Everything is generated locally, so pasting sensitive questions is safe.
Comparing the three scaffold styles in practice
The three styles suit different tasks and output requirements:
Step-by-step
Best for: arithmetic, multi-step logic, structured analysis.
A typical output looks like:
Step 1: Identify the relevant constraints.
Step 2: Apply constraint A to narrow the options.
Step 3: Check the remaining options against constraint B.
Final answer: [conclusion]The numbered structure forces completeness — the model is less likely to skip steps that would expose a flaw in the reasoning.
Think-aloud
Best for: creative analysis, ethical reasoning, open-ended problems.
This style produces flowing prose reasoning rather than a numbered list. It can surface considerations that a rigid step structure would miss, and it reads naturally in a chat context. The downside is that the final answer is sometimes buried in the narrative — less useful when you need to parse the output programmatically.
Scratchpad
Best for: tasks where you want the reasoning available but the deliverable clean, such as code generation with accompanying explanation, or structured data extraction.
The output is split into two delimited sections:
[THINKING]
... reasoning here ...
[/THINKING]
[ANSWER]
... clean final result ...
[/ANSWER]You can extract the answer block programmatically with a simple string split, making this the best style for API use cases where the calling code needs to consume only the result.
When explicit CoT instructions are unnecessary
Newer reasoning-optimized models (those with extended thinking or native chain-of-thought modes) reason internally before producing output. Adding explicit CoT scaffolding to these models is redundant and can actually reduce quality by constraining how the model structures its reasoning. For these models, a direct, well-specified task description is usually more effective than a reasoning template.