Prompt Guide for GPT-4.1
“GPT-4.1 has been trained to follow instructions much more accurately and literally than previous models, which tended to interpret user intent and system prompts more freely.”
With those words, OpenAI introduces a new generation of language models. GPT-4.1 and its smaller variants – mini and nano – are smarter, faster, and more specific. But before you head over to ChatGPT: GPT-4.1 is currently not available in the regular ChatGPT app or via a Plus or Team subscription. Only developers working with the OpenAI API can use the model. And at this moment, it is not yet available in all countries.
Nevertheless, it is certainly worth looking into GPT-4.1 already. Why? Because the model works fundamentally differently than previous versions like GPT-4o. It follows instructions more literally, processes context much more effectively, and is capable of analyzing up to 1 million tokens – which is over 1,500 pages of text in one go. For comparison: GPT-4o stops at approximately 125,000 tokens (190 pages of text). Additionally, GPT-4.1 has a more recent knowledge base up to June 2024 (compared to October 2023 for GPT-4o), and scores 54.6% on SWE-bench Verified, a benchmark for software bug solutions that demonstrates how well the model functions as an independently operating ‘agent’.
However: you can only harness this power if you adapt your prompts to the new way of working. OpenAI has published an extensive prompt guide for this, similar to the document Google released earlier this year. In this blog, we summarize that new approach. Not technically, but practically. And importantly: most tips are also valuable for those still working with GPT-4o or other models.
Step 1 – Role and Goal: what should the language model (such as ChatGPT) do?
Clearly define the expert role the model should assume and what you expect from the model. This helps to establish the context immediately.
Example:
You are a law partner with more than 30 years of experience, specialized in employment law and more specifically in summary dismissal. Your goal is to analyze, based on the information shared with you below, whether an employee of company X can be rightfully dismissed on the spot.
Why this works: by assigning a clear role (employment law partner) and stating the goal (analysis of rightful summary dismissal), the model knows its responsibility.
Step 2 – Instructions: what should the model do and not do?
Provide rules of conduct: how should the model answer, what tone should it use, and what should be avoided?
Example:
● Provide only legally sound answers.
● Mention only risks arising from the shared information, not from general assumptions.
● Use bullet points for each part of the information.
● If information is missing, say: “Based on this document, no statement can be made regarding this.”
Why this works: GPT-4.1 follows instructions literally. If you say ‘do not make assumptions’, the model will comply.
Step 3 – Sub-instructions: refine behavior for specific situations
Here you can provide extra rules for nuance or tone. Think of forbidden topics, example sentences, or when the model should ask a follow-up question.
Example:
● Never use the phrasing “I think that…”. Instead, say: “Based on document Y, it appears that…”
● Ask the question: “Could you indicate which parts you would like further explained?” if the document is too general.
Why this works: sub-instructions make the model more precise and prevent unwanted output (such as suggestive language or legally premature conclusions).
Step 4 – Step-by-step reasoning: think before you answer
Ask the model to plan first and only then provide an answer. Especially useful for complex questions.
Example:
Think about this assignment step by step first.
- Read the document.
- Determine if there are parts of the document with risks for the employer regarding this summary dismissal.
- Summarize for each risk clause why it is risky.
- Provide a summary advice.
Why this works: GPT-4.1 does not have an ‘inner voice’, but it can simulate a reasoning step. By explicitly asking for this, you force the model to be more thorough.
Step 5 – Output format: enforce structure
Specify what the answer should look like, from headings to formatting. This prevents incoherent or messy output.
Example:
Provide the answer in the following format:
● Risks per document: a list of bullets.
● Summary: three sentences with the most important findings.
● Follow-up action: a concrete recommendation.
Why this works: structure is important. If the model knows how you want to see the answer, the responses become much more useful.
Step 6 – Examples: show what you mean
Nothing helps GPT-4.1 better than an example of what you consider ‘good’.
Example of input, where you ask the model:
“Assess whether the summary dismissal in the case below was given with sufficient immediacy.”
Providing an example or examples
Then provide a similar previous case in which the immediacy was legally correctly substantiated. Also add an example of the associated documentation (such as a meeting report or dismissal letter).
What you put in your prompt: “Use this previous case as a reference. Formulate the assessment of the immediacy and the associated documentation in the same way: professional, factual, and legally well-substantiated. Pay particular attention to carefully naming the timeline and the decision moments.”
Why this works: GPT-4.1 learns quickly from examples. The more clearly you show what is good, the greater the chance the model will adopt your style – including tone, structure, and substantive sharpness.
Step 7 – Repeat instructions at the bottom
GPT-4.1 processes long prompts better than ever (up to 1 million tokens), but often remembers instructions at the end better than at the beginning.
Example:
Don’t forget:
● Provide only legally sound answers.
● If something is not in the document, say so.
● Always use the requested format.
Why this works: with long context, the model may ‘forget’ what you said at the beginning. By repeating your assignments at the bottom, you maintain more control.
Bonus: practical tips
Finally, some practical lessons for prompting with GPT-4.1:
● Place instructions at the beginning and the end for long prompts. GPT-4.1 often remembers information at the end better than at the beginning, especially with long contexts.
● Use markdown or XML to make your prompt readable. Use headings (such as ## Instructions) and lists (- point 1) to provide structure. This prevents the model from losing track.
● Work with bullets instead of long blocks of text. This makes it easier for GPT-4.1 to correctly understand and process individual components.
● Let GPT-4.1 work independently until the task is truly finished. You can ask the model to perform a task in steps: first plan, then execute, and then check if everything is correct. For example, when analyzing multiple legal documents: explicitly ask the model to stop only when all parts have been reviewed and processed.
● Check for conflicting instructions. Many errors occur because instructions unintentionally contradict each other. GPT-4.1 then often follows the last instruction. So make your prompt unambiguous and test it step by step.
● Prompt engineering = document design. Don’t think of it as ‘just typing something in’, but as carefully building an instruction document. A good prompt is logically structured, contains clear boundaries, and examples.
● If something doesn’t work: split your prompt. Break complex tasks into smaller pieces or use intermediate steps. This prevents the model from getting entangled in too much input at once.
Conclusion
GPT-4.1 shows that prompting is no longer guesswork, but a skill you can develop purposefully. While previous AI models still filled in a lot themselves, GPT-4.1 listens extremely precisely to your instructions. Those who provide the right role, clearly formulate what the model may and may not do, enforce structure in the answer, and work with concrete examples, will get much more out of the tool. With this approach, GPT-4.1 turns into a reliable sparring partner.