Give a model the same release notes with “summarize this,” and it may produce a broad overview. Change the request to “write for someone uploading a file; explain only how this update affects them,” and the focus shifts to whether an upload can continue after a lost connection. The material and the model stay the same, yet a few words change how the material is handled. Understanding why this can happen is the most interesting part of prompting.
Predicting text teaches relationships behind it
During pretraining, a language model repeatedly predicts what comes next from the preceding content. The objective looks narrow, but the material opens onto a broad world. Continuing a program calls for relationships between variables and operations; continuing a solution calls for following its conditions; continuing a debate calls for identifying what the participants disagree about. Text is full of traces left by people doing things, and prediction gives the model an opportunity to learn their patterns.
For example, remembering “a falling cup can break” helps complete one sentence. Learning relationships involving falling objects and breakable materials helps with many different sentences. Similar patterns recur across contexts, making shared internal representations useful beyond memorizing individual passages. The model gains knowledge that can transfer and combine. Its behavior can feel human-like when it applies familiar relationships to a problem it has never encountered in exactly that form.
Training material also contains many ways of working. Product announcements select changes users care about; technical documentation spells out conditions and interfaces; reviews examine evidence; tutorials unpack intermediate steps. They may discuss the same subject while choosing different information, ordering, and standards of judgment. The model learns overlapping behavioral possibilities, with the current context helping determine how to proceed.
A prompt changes the conditions of this computation
“Write for someone uploading a file” adds the reader’s situation to the context. “Explain only how this update affects them” supplies a criterion for selecting information. The model uses both the release notes and these requirements as it generates text. Different requirements change the probability distribution over possible continuations. This usually leaves model weights untouched; the same parameters perform a different computation on a different input.
The example below places the same material under three prompts. As you switch, notice which facts remain, which details disappear, and who the sentences are meant to help act. The responses are written illustrations, showing the relationship a prompt aims to influence: the model’s interpretation of the task appears in its selection of information.
Summarize these release notes.
The update adds resumable uploads, retries failed chunks up to three times, and refreshes the settings layout.
Cover the changes in the source.
Write for someone uploading a file. Explain how this update affects them.
If your connection drops, your upload can resume without starting over. Failed chunks are retried up to three times.
Select changes relevant to this reader.
Use this pattern: Change → Affected people → Action. Example: Dark mode → Night readers → Select it in Settings. Apply it to resumable uploads.
Resumable uploads → People uploading over unstable connections → Restore the connection and resume the upload.
A concrete example establishes the relationship and output structure.
Examples can be more useful than adding adjectives. “Concise, accurate, professional” leaves considerable room for interpretation; “change → affected people → action needed” makes the intended relationship visible. The early GPT-3 few-shot experiments demonstrated adaptation to varied tasks through examples in the input, without parameter updates. This in-context learning makes a conversation a place to establish temporary rules for a task.
Some prompts also change the process used to complete a task. Asking for comparison criteria before comparing the options puts intermediate results into the subsequent context. The model can continue computing over material it has already developed, reducing the relationships it must handle at once. This helps explain why decomposition can be useful: productive intermediate states supply new input for the next step. Repeating the same instruction adds no such state.
When writing a prompt, I care about making the task recognizable to the model. Who is the reader? What should the answer help them do? What material counts as evidence? What counts as completion? These facts describe the work more clearly than “you are a world-class expert.” A role can help by bringing relevant standards of judgment into play. Stating those standards directly also reduces the need to infer them from a title.
Training remains the basis for this behavior. Pretraining supplies knowledge and patterns of behavior, while later instruction training makes the model more accustomed to treating a user’s request as a task to fulfill. As that habit strengthens, a short request can elicit behavior that once needed elaborate explanation. Prompt length is therefore a poor goal on its own. The aim is to give the current model enough information to organize its learned capabilities around the work.
This leads to a natural next step. The model can explain how an upload resumes after a lost connection, but finding where a particular upload has stopped requires looking at its environment. A prompt can direct the model toward a useful way of thinking. Give it ways to observe and act, and that thinking can connect to events as they unfold.