Three Prompting Habits OpenAI’s Guide Warns Against

Prompt quality often decides whether a GPT model returns something useful or something generic. OpenAI’s own prompt engineering guide lays out how its current models respond to instructions, and several of its recommendations point to habits that quietly weaken results. The guidance is published at https://developers.openai.com/api/docs/guides/prompt-engineering and applies to the GPT-5 series of models.

Treating The Model Like It Can Read Intent

The guide describes a current GPT model as comparable to a junior coworker who needs explicit instructions. It notes that these models benefit from precise instructions that explicitly provide the logic and data required. The implication is that vague requests place too much on the model to infer.

When a prompt leaves the goal implicit, the model fills the gap with its own assumptions. Spelling out the desired logic, the constraints, and the expected form of the answer removes that guesswork. This is less about writing longer prompts and more about being direct regarding what the response should contain.

The guide also points to structure as a way to make intent clear. It describes message formatting with Markdown and XML, organizing a prompt into sections such as Identity, Instructions, Examples, and Context. Separating instructions from supporting material helps the model distinguish what it is being asked to do from the data it is being asked to use.

Leaving Out The Context The Model Needs

A second habit the guide addresses is omitting information the model cannot already know. Under its section on including relevant context information, the guide explains two purposes for adding context: supplying proprietary data the model was not trained on, and constraining responses to a beneficial set of resources. It connects this practice to retrieval-augmented generation, where relevant material is added to the prompt so the model can ground its answer in it.

The same section introduces a counterweight. The guide discusses planning for the context window, describing the model’s memory as a limit measured in tokens. Context is necessary, but it competes for finite space. The practical reading is to include the material the model actually needs for the task and to be deliberate about it, rather than padding a prompt with everything on hand. Relevant context helps; filler consumes the budget that the answer itself needs.

Handing Over A Complex Task Undivided

The third habit involves asking for too much at once without giving the model a way to manage it. In its discussion of agentic tasks, the guide recommends planning and persistence. One sample instruction tells the model to decompose the user’s query into all required sub-requests and confirm that each is completed.

That advice can be applied before the prompt is even sent. Breaking a large request into smaller, ordered subtasks gives the model a clearer target for each step. The guide also describes asking the model to plan extensively before acting and to reflect on outcomes, which works best when the overall task is divided into pieces that can be reasoned about individually.

Applying The Guidance

The common thread across these habits is clarity. The guide’s other tactics reinforce it: few-shot learning, where input and output examples let the model pick up a pattern, and message roles such as developer, user, and assistant, which carry differing levels of authority over the model’s behavior. Being specific, supplying only the context that matters, and splitting complex work into parts are habits worth building, and OpenAI’s documentation is the most direct reference for how its models expect to be prompted.