How Can a Prompt Elicit a Model's Capabilities?

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. The question is why a prompt can change how the material is handled.

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

Being able to continue many kinds of text does not by itself establish a habit of following user requests. Instruction following also has to be learned. InstructGPT fine-tuned a pretrained model on human demonstrations, then trained further using human evaluations of responses. This made pretrained knowledge and relationships easier to organize into requested responses. An everyday prompt calls on a model that has already learned something about accepting tasks.

“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.

What do common prompting techniques actually change?

Once a prompt is understood as supplying conditions for a particular computation, familiar techniques fit into a common picture. Some narrow the interpretation of the task; some demonstrate a relationship to reproduce; others arrange intermediate work or create more opportunities to find a good answer. A technique earns its place when the condition it supplies is what the task actually lacks.

• Roles and goals supply a standpoint for judgment. “You are a product manager” suggests familiar priorities, but product managers also write requirements, rank work, and present reports. Adding “the reader has just had an upload fail and needs to decide whether to retry or wait” narrows those possibilities to the present job. A role can be shorthand for an audience, responsibility, and criteria, but those meanings need to be clear.

• Examples make vague standards recognizable. “Concise, accurate, professional” still permits many outputs. An example connecting “what changed → who is affected → what to do next” demonstrates the desired relationship. GPT-3’s few-shot experiments showed adaptation to tasks through examples in the input, without parameter updates. Examples are particularly useful for style and boundaries that are difficult to enumerate: a good sample alongside a seemingly good one that misses an essential condition can communicate more than ten adjectives.

• Constraints and delimiters assign different jobs to different content. Separating background, source material, and deliverable requirements reduces the chance of treating source text as an instruction or an example as a fact. Tags, headings, and code fences can all serve this purpose; no particular symbols are inherently smarter. Requesting JSON constrains the delivery interface. When another program depends on it, use supported structured-output constraints and validate the result as well. A boundary written in a prompt is, first of all, a boundary for the model to understand.

• Decomposition and intermediate artifacts give later computation something to work with. Establishing dimensions before comparing two proposals lets subsequent judgments use them one by one. Checking a fact sheet before writing a report gives the prose a stable basis. The benefit comes from preserving relationships that later work needs. Splitting a task into ten steps without adding useful information creates nine extra handoffs and may lock in a route that should change as discoveries arrive.

Why did “think step by step” once matter so much?

Early models were often asked to jump directly from a question to an answer. Putting worked reasoning examples in the input made a response with intermediate steps more likely; those generated steps then supplied conditions for subsequent computation. Chain-of-thought prompting demonstrated how language could carry intermediate computation. The instruction helped organize a way to unfold and reuse several relationships during problem solving.

Other techniques spend effort on selection. Generating several proposals and comparing them, or voting across different solutions, broadens the search. Self-consistency improves performance by sampling reasoning paths and aggregating their answers. Reflection follows a related idea, but “check again” merely creates another occasion for judgment. A counterexample, an executed test, or a check against the source adds new evidence. Five answers that share a misunderstanding can all vote for the same error.

Today, many reasoning models have learned to allocate intermediate computation during training, while agent models have learned to retrieve information, use tools, and respond to failure. Behaviors that once needed prompting have become defaults. OpenAI’s reasoning-model guidance consequently recommends direct, clear requests and notes that instructions to think step by step may be unhelpful or harmful. The principle of supporting difficult judgments with intermediate work remains useful; how much of that work a person must prescribe needs reassessment.

What should a prompt make clear today?

As the excitement around prompt engineering recedes, a practical job remains: connecting general capability to a particular person’s intentions. Models increasingly know how work is usually done, but they still need to know which result matters to you, what tradeoffs you accept, and how much authority to exercise. Current provider guidance increasingly addresses such collaboration conditions, including initiative and instruction priorities and adjusting guidance to the particular model. Strong reminders can produce excessive action when carried over to a model that is already proactive.

1. The outcome and tradeoffs. Who will use the deliverable, for what purpose, and which requirement wins when demands conflict? For an essay, “help readers understand why this works” and “cover every topic” lead to very different writing.

2. Facts and authority. Which materials are current evidence, and which are merely examples? Where may the model investigate, what may it change independently, and which decisions belong to the person? When information changes, access to a source and a requirement to check it are more useful than an old answer permanently embedded in the prompt.

3. Completion and delivery. Must code be verified after it is written? Which file should contain the article? What evidence belongs in the final report? Clear conditions let the model choose a process. Prescribe a process when that process itself affects quality—for example, composing the sentences of a paragraph before writing that paragraph to a file.

For the upload announcement, that could look like this:

Use these release notes to draft a help-page paragraph for users whose uploads failed. Readers should be able to decide whether to retry or wait. Keep conditions that affect that decision; consult the relevant documentation when information is missing rather than guessing product behavior. Match the existing help-page style. Check that every product claim is supported by the materials, and report where the draft is saved.

This prompt leaves the individual actions open while establishing selection criteria, evidence, style, an artifact, and a completion standard. A simpler task may need less. When a model repeatedly misunderstands a particular point, add an example targeted at that point.

This gives a way to reassess older techniques. Clear goals, relevant examples, separated sources, and useful intermediate artifacts retain their value. Repeated status claims, promised rewards, and mandatory long derivations for every task do not deserve to survive as default rituals. Long prompts can be useful when their length comes from real task conditions. Remove reminders that the model has already internalized, that have no bearing on this job, or that conflict with one another.

Finally, judge prompts through their tasks. Compare revisions on realistic examples, including boundary cases that are easy to misunderstand, and revisit old reminders after a model upgrade. One impressive response does not establish a template’s value, and one failed instruction does not establish a lack of capability. Distinguish an unclear goal, insufficient information, and a task the model cannot yet perform. The first two can often be improved through input. When that information comes from a changing world, prompting connects to the context engineering explored next.