To prompt a reasoning model, keep the brief simple and direct: state the goal, the constraints and what a successful answer looks like. Do not add a “think step by step” cue as a default. In Reasoning best practices, OpenAI says these models already perform reasoning internally, and that prompts such as “think step by step” or “explain your reasoning” may not help and can sometimes hinder performance. The same guide says to try a zero-shot prompt first, use delimiters to mark the parts of the input, and be specific about the end goal.
That is a different habit from chain-of-thought prompting, which asks a general chat model to show its intermediate steps. The earlier article already notes the caveat. This one is the how-to for models that think privately first, then write.
For eBook and workbook projects, Interactive eBook Builder is a $59 one-time Notion AI template with Spec Builder, Drafter, Polisher and SEO BOT already set up. The Writers Workshop includes a Prompt Engineering Lab where you can change one instruction, predict what will happen, and test the brief on a normal case and a failure case. Notion AI access and usage are separate from the purchase. The examples below are teaching prompts written for this article. They are not paid Builder prompts, and they are not recorded model outputs.
What is a reasoning model, and how is it different?
OpenAI splits its models into two families. In the same best-practices guide, it describes o-series models as “the planners”: they are trained to think longer about complex, ambiguous work, then decide. It describes GPT models as “the workhorses”: they are faster and cheaper, and they are built for straightforward, well-defined tasks. The guide is explicit that one family is not simply better than the other. Many workflows use both: a reasoning model to plan, and a GPT model to execute a defined writing job.
Other vendors now ship their own thinking or extended-thinking modes. The labels change, and the advice in a chat app can differ from the API docs. The practical test is the same. If the model already reasons internally before it answers, treat it as a reasoning model: give it a clear problem, not a script of how to think. If it does not, a written chain of thought can still help on planning tasks. Check the current documentation for the model you are actually using.
OpenAI also notes an API-specific detail: starting with o1-2024-12-17, reasoning models use developer messages rather than system messages. In ChatGPT, you will not usually see that label. The idea that still matters for writers is simpler. Put standing rules and the task in clearly marked sections, and do not bury the request in a long speech about method.
How should you prompt a reasoning model?
OpenAI’s current list is short, and it is worth following closely:
- Keep prompts simple and direct. The models “excel at understanding and responding to brief, clear instructions.”
- Avoid chain-of-thought prompts. Because reasoning happens internally, “think step by step” or “explain your reasoning” is unnecessary.
- Use delimiters. Markdown, XML tags and section titles help the model tell notes, rules and the request apart. That is the same discipline described in How to Use Delimiters in Prompts.
- Try zero-shot first, then few-shot if needed. Reasoning models often do not need examples. If you add them, they should match the instructions closely, because a mismatch can produce poor results.
- Spell out constraints and the end goal. If a chapter must stay under 400 words, or a claim must come only from your notes, say so. OpenAI also suggests encouraging the model to keep reasoning until the answer meets those criteria.
This is why a familiar “good” prompt can disappoint on a reasoning model. A brief that worked on a GPT model by walking through five planning steps may now get in the way. You are no longer teaching the model how to think. You are defining the problem well enough that its internal search can do that work.
When should an eBook writer use a reasoning model?
OpenAI’s guide says to reach for a reasoning model when the job is complex, multi-step, or ambiguous, and when accuracy matters more than speed or cost. It lists patterns such as navigating incomplete information, pulling the relevant facts out of a large pile of unstructured text, and multi-step planning. It recommends GPT models when the task is well defined and speed or cost matter more.
That maps reasonably onto book work:
| Job | Better default | Why |
|---|---|---|
| Choosing a chapter angle from messy, conflicting notes | Reasoning model | The notes are incomplete and some items conflict |
| Checking whether a draft is supported by your source list | Reasoning model | The model has to compare claims against evidence |
| Deciding what to cut from an overloaded outline | Reasoning model | There is a real trade-off, not a style polish |
| Drafting a section from an approved outline | GPT or a standard chat model | The decisions are already made |
| Turning a settled section into a workbook line | GPT or a standard chat model | The job is narrow and well defined |
| Tightening a paragraph for clarity | GPT or a standard chat model | Extra thinking time is usually wasted |
These are starting points, not guarantees. A reasoning model can still invent a tidy plan that ignores a constraint you forgot to write down. A standard model can still handle a planning pass if you ask it to show its steps. Test one chapter both ways if the job matters.
Example 1: plan a chapter without scripting the thinking
A prompt written for a general chat model often looks like this:
You are going to plan a chapter for new freelance designers about
naming a floor rate. Think step by step. First list every theme in
the notes. Then rank the themes. Then resolve conflicts. Then write
a five-heading outline. Then explain why you chose that order. Then
write the chapter.
Notes:
- The reader currently charges favors or student-level prices.
- I want them to set a floor before the next inquiry, not a full menu.
- I mentioned a $75 example in one note and a $120 example in another.
- I do not have survey data, and I do not want invented averages.
That brief tells the model how to think, asks it to explain the thinking, and then asks it to write the chapter in the same turn. On a reasoning model, OpenAI’s guidance points the other way: state the goal and the success criteria, and leave the method alone.
Goal: Plan one chapter of a short eBook for new freelance designers.
The chapter should help the reader name a floor rate before their
next inquiry, not build a full pricing menu.
Notes:
- The reader currently charges favors or student-level prices.
- I mentioned a $75 example in one note and a $120 example in another.
- I do not have survey data, and I do not want invented averages.
Constraints:
- Use only these notes. If something is missing, list it under
"Need from me" instead of guessing.
- Do not invent industry averages, quotes or studies.
- Do not write the chapter yet.
Success looks like:
- A short "Decisions and gaps" list
- A five-heading outline that a later drafting pass could follow
- One workbook line the reader will complete
- Any unresolved conflict (such as $75 vs $120) flagged, not split
the difference
The second brief is shorter, but it is stricter. It names the reader, the job, the limits and the shape of a good answer. It also stops at the plan, which keeps you from accepting a draft before the conflicts are settled.
Example 2: check a draft against your notes
Reasoning models are a reasonable fit for a “needle in a haystack” pass: you paste a draft and a short source list, and you want mismatches, not a rewrite. A weak prompt asks for a general critique and a better chapter. A tighter one defines the test.
Goal: Check this draft against the notes I supplied. I need to know
what is supported, what is missing and what was added.
Notes:
- The only fee mentioned in my material is a $90 floor for a one-page
logo revision.
- I have no client survey and no "industry average."
- The chapter should end with a fill-in line for the reader's floor.
Draft:
[paste the section]
Constraints:
- Do not rewrite the draft in this pass.
- Do not add sources I did not provide.
- If a claim is not in the notes, label it "unsupported."
Success looks like:
- A table with three columns: Claim, Supported / Unsupported / Missing,
and Evidence from the notes (or "none")
- A short list of invented numbers, quotes or averages, if any
- No suggested rewrite until I ask for one
This is the same spirit as a later editing pass: inspect first, change second. If you want the model to revise, do that in a separate request after you have accepted the findings.
Example 3: do not micromanage a well-defined writing job
When the outline is already approved, a reasoning model is often more than you need. This brief is enough for a standard chat model, and it is also a good test of whether you are over-prompting:
Write the "Name the floor before the call" section from this outline.
Reader: new freelance designer who undercharges.
Length: 280 to 320 words.
Tone: calm, direct, no hype.
Use only the outline and the notes below. If a figure is not in the
notes, do not invent one.
Outline:
- Why a floor is different from a full pricing menu
- A worked example using the $90 floor already in my notes
- One mistake: quoting a lower number live and backing yourself into it
- Close with a fill-in line for the reader's floor
Notes:
- The only figure I am willing to publish is a $90 floor for a one-page
logo revision.
If you are using a reasoning model here, you can still use this brief. You do not need to add “take your time” or “explain your reasoning” unless you are following a vendor’s current, specific instruction to do so. Keep the job definition; drop the method talk.
How should you test the same brief on two model types?
The Prompt Engineering Lab in the Writers Workshop is built for a small, controlled comparison: change one instruction, predict the result, and run the brief on a normal case and a failure case. A useful test for this topic is:
- Take the planning brief from Example 1.
- Run it once on a reasoning model, as written.
- Run the same brief on a standard chat model, then run a second version that adds “think step by step” and asks to see the work.
- Compare the plans, not the prose style. Did the model flag the $75 versus $120 conflict, or did it invent a compromise? Did it stay inside the notes?
Spec Builder is the stage where a planning brief like this belongs. Drafter should receive an approved outline, not a pile of unresolved conflicts. Polisher can use a later, narrower brief to check a finished section against the same notes. None of those stages needs the model’s private reasoning printed in the book. The chain is a working aid for you.
What should you watch for when the answer looks confident?
A fluent plan is not automatically a faithful one. The model can still smooth over a gap, pick one of your conflicting figures without telling you, or treat a “need from me” item as if you had already decided it. Read the output against your notes, especially numbers, quotes and claims that sound like research.
Also remember that vendor guidance changes. OpenAI’s page is about its reasoning models, and it includes API details that do not appear in every chat interface. If you are using a thinking mode from another vendor, read that vendor’s current prompting notes before you copy these examples wholesale.
FAQ
Should I tell o1 or o3 to think step by step?
Usually no. OpenAI’s reasoning best practices say those models already reason internally, and that “think step by step” or “explain your reasoning” is unnecessary and can sometimes hinder performance. Give a clear goal, constraints and success criteria instead. If you are not sure whether the model in front of you is a reasoning model, check its documentation, then test one planning task with and without the cue.
Do few-shot examples help reasoning models?
Not as a default. OpenAI recommends trying a zero-shot prompt first, because reasoning models often do not need examples. Add a few only when you have a precise output shape to show, and make sure the examples match the instructions. Conflicting examples can hurt more here than they do on a standard model. That is the inverse of the usual zero-shot versus few-shot advice for general chat models, where examples are often the first upgrade.
Can I use the same planning prompt I already wrote for ChatGPT?
You can start from it, then cut the method. Remove “think step by step,” remove requests for a long explanation of the reasoning, and keep the reader, the goal, the constraints and the definition of done. If the old prompt also asked the model to plan and write in one turn, split those jobs. A reasoning model can still do both, but you will make better decisions if you approve the plan first.
Before you hit send
- Is this a planning or checking job, or a well-defined draft from an approved outline?
- Have you stated the goal, the constraints and what success looks like?
- Did you remove “think step by step” and other method cues that the current docs advise against?
- Are your notes, rules and request in clearly marked sections?
- If you added examples, do they match the instructions exactly?
- Will you inspect the plan or the fact-check before anyone drafts the chapter?
A reasoning model is not a reason to prompt less carefully. It is a reason to prompt more like a brief and less like a tutorial.