Four Rules for Prompting Reasoning Models
Give reasoning models clear structure without making them overthink
- Difficulty
- Easy
- Time to result
- ~days to results
- Steps
- 4
- Confidence
- 98%
This framework adapts prompting to models that can plan and reason internally. Instead of prescribing every intermediate thought, give the model a brief, unambiguous objective and let it determine the sequence of work. Organize separate instructions, evidence, and output requirements with XML-style delimiters or descriptive section headings so the model can distinguish their roles. When retrieval is involved, filter the available material and supply only the documents or passages most relevant to the task. The mechanism is constraint through clarity rather than control through verbosity: a clean objective focuses the model, structural boundaries prevent input categories from blending together, and selective context reduces irrelevant reasoning paths. The result is a shorter prompt that still supports sophisticated multi-step work without inviting unnecessary complication.
Origin
Extracted from Marketing Against the Grain during Kieran Flanagan's explanation of OpenAI's four recommendations for prompting the o1 reasoning-model family.
Core principles
- 01Trust the model to perform its own reasoning
- 02Prefer brief and direct instructions
- 03Separate distinct inputs and requests clearly
- 04Provide only context that materially affects the answer
How to run it
- 1
State the task directly
Describe the desired outcome with brief, clear instructions. Remove background commentary that does not change what the model must produce.
Pro tip Write the first version as though assigning a task to a capable colleague in Slack.
Watch out Do not confuse simplicity with vagueness; retain concrete success criteria.
- 2
Remove chain-of-thought requests
Ask for the result rather than instructing the model to think step by step or expose its reasoning. Allow the reasoning model to determine and execute the intermediate sequence internally.
Pro tip Request a concise conclusion or decision rationale when useful, not a transcript of hidden reasoning.
Watch out Adding elaborate reasoning rituals can be redundant and may complicate the response.
- 3
Delineate the inputs
Use XML tags, section titles, or similarly clear boundaries to separate the task, evidence, constraints, and output format. This helps the model identify the purpose of each part.
Pro tip Use consistent labels such as <task>, <context>, and <output_requirements>.
Watch out Do not add structural markup that creates more complexity than the underlying request.
- 4
Limit retrieved context
Select only the documents and passages that are relevant to the requested outcome. Exclude loosely related material that could pull the model toward unnecessary interpretations.
Pro tip Filter retrieval results before inserting them into the final prompt.
Watch out Over-filtering can remove evidence needed for an accurate or properly qualified answer.
In the wild
A founder asks a reasoning model to compare five acquisition candidates. The prompt states the decision to be made, separates financial data from strategic criteria with labeled sections, and includes only the relevant records for each company. It asks for a ranked recommendation and key risks without demanding a step-by-step thought process.
→ The founder receives a focused comparison that can be used much earlier in the acquisition process.
A marketer supplies a concise description of a lightweight internal application, clearly labels the required features and output constraints, and omits unrelated company documentation. The reasoning model determines the implementation sequence rather than being given a long procedural prompt.
→ The model can produce a coherent multi-step implementation from one structured request.
Common mistakes
Writing an enormous prompt
Long prompts can bury the actual objective beneath instructions and background material that do not improve the result.
Forcing visible step-by-step reasoning
Reasoning models already perform internal reasoning, so demanding a chain-of-thought process adds unnecessary instruction.
Dumping every available document
Unfiltered retrieval gives the model irrelevant context and can cause it to overcomplicate its response.
Is it for you?
Best for
It is best for people using advanced reasoning models on coding, analysis, research, and other multi-step tasks.
Not ideal for
It is not ideal when a task genuinely requires extensive reference material that cannot be filtered without losing essential evidence.
From the transcript
“keep prompts simple and direct”
“avoid the Chain of Thought prompts since these models perform reasoning internally”
“only give it the most relevant information because you don't want you don't want the model to over complicate its response”
From the episode
Chat GPT-o1 is Mindblowing!! Everything You Need To Know