Overview

Create, edit, test, version, evaluate, improve, and deploy prompt behavior from the project workspace

Prompts are the main levers your AI application uses at runtime. A prompt can include model settings, messages, variables, files, tools, response formats, evaluators, datasets, versions, and deployments.

In the self-improving loop, prompts are the change surface. Logs and Behaviors explain how the current prompt behaves in production; Evaluators and Datasets define what must be preserved; Improve or human editing creates the next candidate; Deploy makes the reviewed version available to your application.

What a prompt contains

Area	What you configure
Model settings	Provider, model, temperature, max tokens, response format, and other generation controls.
Messages	System, user, assistant, and tool messages with text, images, PDFs, and examples.
Variables	Runtime inputs such as {{user_question}}, {{context}}, or API/prompt-backed values.
Tools	Project tools and optional MCP server tools when enabled.
Evaluation	Linked datasets and evaluators for testing across cases.
Versioning	Drafts, snapshots, deployment history, and Improve candidates.

Build and test

The authoring loop is:

1. Choose a model and configure generation parameters.
2. Compose messages with roles and content.
3. Add variables for runtime inputs.
4. Attach tools when the model needs external actions or data.
5. Run the prompt in the playground with representative inputs.
6. Link datasets and evaluators before deployment.

For the step-by-step workflow, see Build and test prompts.

Messages, variables, and tools

Older prompt docs called this the Iterate workflow. In the current Platform sidebar, the same authoring concepts live under Prompts.

Concept	Use it for
Roles	Structure system, user, assistant, and tool context.
Text	Write instructions, examples, comments, and variable placeholders.
Images and PDFs	Add multimodal context for supported models.
Variables	Make prompts reusable across users, datasets, API inputs, and chained prompts.
API variables	Fetch live data into prompt context.
Prompt variables	Use one prompt’s output as another prompt’s input.
Tools and MCP	Let the model call external functions, APIs, retrieval systems, or MCP tools.

Keep variable names stable. Datasets, dynamic columns, API integrations, and chained prompts depend on them resolving consistently.

Playground

Use the playground to execute prompts with specific inputs, inspect model responses, compare configurations, and debug tool calls before running larger evaluations.

The playground is useful for fast iteration, but it is not a replacement for datasets and evaluators. Use it to understand behavior, then run evaluations to test many cases.

Versioning and deployment

Editing a prompt changes the draft. Deployment creates a snapshot for an environment. Your application should read from the deployment environment it expects, not from an arbitrary draft.

State	Meaning
Draft/editor state	Where prompt editing and playground work happen.
Prompt version or snapshot	Captured state used for review or deployment.
Deployment snapshot	What an application environment reads at runtime.
Improve approval	Can apply a candidate and deploy it, while Edit & approve applies it without deployment.

Use prompt versions and deployment snapshots as the release model for production changes.

Prompts in the Platform workflow

• Tools define callable functions or HTTP-backed actions.
• Evaluators define success criteria.
• Datasets provide test cases.
• Logs show deployed prompt behavior in production.
• Behaviors cluster recurring outcomes.
• Improve proposes prompt changes from those signals.
• Deploy ships a prompt version to an environment.

Build and test prompts

Configure messages, variables, model settings, tools, datasets, and playground runs.

Version and deploy

Understand drafts, snapshots, deployment environments, approval, and rollback.

Tools

Define reusable tool schemas and HTTP-backed actions.

Evaluators

Define pass/fail, scoring, cost, latency, and formatting criteria.