Prompt Studio: what goes where, and why

Prompt Studio improves the instructions you send to a model. Start with a clearly defined task, separate the instructions from the changing inputs, and review the rewritten prompts before using them.

1. Describe the task in one sentence

The task description tells the optimizer what you are trying to accomplish.

Generate search marketing keywords for healthcare services.

Keep it brief. Put the detailed requirements in the system or user prompt, where the model performing the task will receive them.

2. Use the system prompt for rules that stay the same

The system prompt defines the model’s role and the requirements that apply across requests. Good candidates include output format, length limits, tone, domain restrictions, and how to handle missing information.

For our keyword task:

You generate search marketing keywords for healthcare clinics.

Return exactly 3 keywords relevant to the requested service.
Each keyword must contain no more than 5 words.
Use a numbered list, with one keyword per line.
Do not include explanations or additional text.

Specific requirements are more useful than “be accurate” or “produce great results.” Describe what a satisfactory answer must contain.

3. Use the user prompt for the current request and its inputs

The user prompt says what to do on this invocation and supplies the information needed to do it.

Suggest keywords for this healthcare service: {{service}}.

Here, the instructions stay the same while the service changes.

Use this distinction when deciding where a requirement belongs:

These are useful defaults, not a requirement to fill both prompt fields. Prompt Studio accepts a system prompt, a user prompt, or both. Avoid conflicting instructions between them.

4. Use {{variables}} for values that change

A double-brace placeholder marks an input your application supplies when using the prompt.

Template:

Suggest keywords for {{service}} in {{city}}.

Inputs for one request:

{
  "service": "physical therapy",
  "city": "Austin"
}

User message sent to the model:

Suggest keywords for physical therapy in Austin.

Use clear names such as {{service}}, {{city}}, or {{customer_message}}. Keep their spelling consistent and supply a value for every variable.

The placeholders survive optimization. Your application must substitute the actual values before sending the resulting messages to the model. Copying the optimized template does not perform that substitution.

For a reusable prompt, keep changing data in variables instead of hardcoding one customer, document, or service into the instructions.

5. Make it a task prompt

A task prompt gives the model an actionable job: what to do, what information to use, and what an acceptable answer looks like.

“Healthcare marketing” names a topic. “Generate exactly three healthcare search keywords for {{service}}, each no longer than five words” defines a task.

The system and user messages together form the prompt-set. Submit the instructions you intend to run, rather than only describing a feature you would like someone to build.

6. Choose the model that will perform the task

The Target task model is the model you intend to use for the actual work. Prompt Studio tunes the rewrite for that model.

The rewriter, also called the reflection model, improves the instructions. These models have different jobs: one rewrites the prompt; the other uses it to answer your requests.

Selecting a target model does not train or change its weights. It changes which model the prompt is being optimized for.

7. Review the optimized prompt-set

The result contains the rewritten messages, their roles, the target model, and the retained variables. You can copy individual messages or the result payload.

Review it before adoption:

• Are the original requirements still correct?
• Are the variable names preserved?
• Has it introduced examples, assumptions, or restrictions you did not intend?
• Does the requested output fit what your application expects?

An optimized prompt may be longer. Clearer instructions can improve consistency or reduce unnecessary output; fewer input tokens are not the only objective.

For zero-shot optimization, the result is a rewrite without a scored task-evaluation loop. It still needs testing.

8. Read Serving cost as an estimate

The Serving cost card estimates the cost of using the optimized prompt on the target task model. It is separate from the cost of running the optimizer.

The current estimate uses a token basis and a blended input/output rate from the live catalog. Actual cost depends on the values substituted into the template and the response produced.

Use it for an initial cost check. Then benchmark the original and optimized prompt-sets on representative inputs, using the same quality criteria, before deciding which version to deploy.

Continue with How much accuracy can you afford? for a worked benchmark comparing quality and serving cost.