Quickstart
This guide takes you from a new account to a correctly structured first project. It does not run paid model work until you explicitly start an optimization or judge.
Choose or create an Organization
An Organization contains the people, billing account, and shared model configuration for a team. Use the organization switcher at the top of the sidebar to select the team that should own the work.
Use a durable company or team name rather than the name of a single experiment. You can invite members later from Organization settings → Members.
Create an AI Application
From Home, choose New AI Application. An AI Application should map to one application, agent, workflow, or production use case with a shared access and usage boundary.
Good examples:
- Customer Support Assistant
- Contract Extraction API
- Product Search Copilot
Avoid creating a new AI Application for every prompt version. Versions and experiments live inside AI Components.
Create an AI Component
Open the AI Application and create a component. Choose the workflow that matches the thing you are measuring:
| Component type | Use it for |
|---|---|
| Prompt Optimization | Classification, text generation, and structured-output prompts |
| Agent Optimization | Complete agents that submit outputs, artifacts, and traces |
| MCP Optimization Preview | Tools discovered from an MCP server |
| Tool Selection Optimization Preview | Tool definitions pasted directly into Promptic |
Component type determines the dashboard workflow and the data attached to the component. Use a
stable functional name such as support-agent or invoice-extractor.
Configure model access
Promptic-managed models are the fastest way to start. If your organization uses its own provider credentials, configure them before starting work under Organization settings → Models or the AI Application’s Settings → Models page.
Read Models and providers before choosing managed access, BYOK, Google Vertex AI, or a custom OpenAI-compatible endpoint.
Complete one measurable workflow
Do not start by configuring every feature. Pick one:
- Send one real trace with the SDK quickstart.
- Create a small prompt experiment with 20–50 representative examples.
- Create a tool-optimization dataset and establish a baseline against representative queries.
Before trusting a result, inspect the underlying examples and confirm the evaluator measures the behavior that matters to your users.
Invite the team
Organization membership provides access to the team. AI Application membership controls access to application data. Add people at the narrowest level they need and review roles before sharing API keys or provider configuration.
See Teams and access.