What is a prompt?
A prompt is the input given to an AI model to produce an output: the instructions, context, examples and question that tell a language model what to do, what information to use, what format to answer in and what to avoid, written in ordinary language rather than code. In business systems, a prompt usually combines fixed instructions written by the organization with the user's message and any documents retrieved for the request.
Prompts are how most organizations steer a model without retraining it. The word covers a one-line question typed into an assistant and a multi-page instruction set governing a production system. Designing, testing and maintaining those instructions is called prompt engineering, and in many deployments it is where most of the system's behavior is actually defined.
How a prompt works
A production prompt is assembled at the moment of each request from several parts, which the model then reads as one continuous text.
| Part | What it does | Example in a customer service setting |
|---|---|---|
| System instructions | Set the role, rules, tone and limits that apply to every turn | Answer only from the policy excerpts provided; never quote an interest rate |
| Context | Supplies the facts needed for this request | The customer's plan and the relevant section of the fee schedule |
| Examples | Show the desired format or reasoning through worked cases | Two sample replies that confirm a change before making it |
| User input | Carries the person's actual message | "Can I move my due date?" |
| Output format | Specifies the structure of the answer | A structured object with a decision field and a reply field |
Because everything arrives as text, the model has no built-in way to tell trusted instructions from untrusted content. An email, web page or document placed in the context can contain sentences that read like instructions. Prompt injection is the name for exploiting that weakness, and it is the main reason security teams treat prompts as part of the attack surface.
Why prompts matter
How a model is prompted can matter as much as which model is used. In the study that introduced chain-of-thought prompting, Wei and colleagues showed a 540-billion-parameter model eight worked examples that reasoned step by step. On a benchmark of grade-school math word problems, its accuracy rose from 17.9% with standard prompting to 56.9%, with no change to the model itself.
The same sensitivity cuts the other way. When much of a system's behavior lives in prose, an innocent edit can change results in ways nobody intended, and two versions that look equivalent can behave differently. Organizations that rely on prompts in production manage them like code: versioned, reviewed, tested against a fixed set of cases before release and monitored afterward.
Where prompts are used
- Assistants for staff and customers. Instructions that set scope, tone and the sources an assistant may use.
- Drafting. Templates for emails, summaries and reports that a person reviews before sending.
- Classification and extraction. Asking a model to label a request or pull details out of a document without training a dedicated model.
- AI agents. System instructions that frame each step of a task, alongside the tools and logic that carry it out.
- Evaluation. Using one model, through a carefully written prompt, to grade another model's outputs.
Rules that must hold every time, such as never quoting a rate before eligibility is checked or always reading a required disclosure, should not live in a prompt alone. They belong in the structure around the model, where they can be enforced and audited rather than requested.
What a good prompt looks like
- Specific instructions. It says exactly what to do and what not to do, with the reasons where they help.
- Separated content. Trusted instructions are kept apart from customer messages and retrieved text, and marked as such.
- Worked examples. A few representative cases show the expected output better than long descriptions.
- Minimal sensitive data. Only the personal information the task needs is included.
- Version control. Every change is recorded, reviewed and reversible.
- A test set. Each version is run against the same cases before it reaches customers.
How to measure a prompt
- Task accuracy. Correct outputs on a fixed evaluation set drawn from real requests.
- Format compliance. The share of outputs that match the required structure.
- Injection resistance. The share of adversarial test inputs that fail to override the instructions.
- Regression rate. Cases that passed before an edit and fail after it.
- Cost and latency. Tokens and response time per request, since longer prompts cost more and run slower.
Compliance considerations for prompts
A prompt is where an organization tells a model how to behave toward customers, so the rules about that behavior apply to what the prompt says. These are the main US considerations; banking, insurance and other regulated fields add their own.
- Instructions about identity. A prompt that tells a model to present itself as a person runs into state bot laws. California prohibits using a bot to mislead someone about its artificial identity in order to encourage a sale, and Utah requires a business using generative AI to disclose it when a consumer asks.
- Security of what a prompt can reach. Prompt injection is an application security risk. Financial companies under the FTC Safeguards Rule must adopt secure development practices for applications they build that handle customer information.
- Customer data placed in prompts. Every detail in a prompt is shared with whoever runs the model. Financial institutions rely on contracts that limit how a provider may use that information, and state privacy laws expect the data shared to be no more than the task requires.
Frequently asked questions
What is a system prompt?
The fixed instructions an application sends with every request, setting the model's role, rules and limits. Users usually never see it.
What is prompt engineering?
The practice of designing, testing and refining prompts so that a model produces reliable results for a given task.
What is prompt injection?
An attack in which text inside a message or document is written to look like instructions, so the model follows it instead of its intended rules. It can expose data or trigger unintended actions.
Can a well-written prompt make a model reliable?
It can make a model much more consistent, but not guaranteed. Anything that must always happen should be enforced by the system around the model, with the prompt as one layer of control.
Is a longer prompt better?
Not necessarily. Longer prompts cost more, run slower and can bury the instructions that matter. Clear, specific and tested beats long.
Related terms
More than a prompt.
Encore AI's agents do not depend on a long prompt to behave.
Scale your best to be the best.
Book a Demo