Prompts and resources
Versioned prompts vs knowledge resources: playground, compare, naming, environments, and pairing them with agents.
Prompts tell the model how to behave. Resources give it what to treat as ground truth for a turn. Together they reduce drift without stuffing everything into one giant instruction block.
How prompts and resources work together
| Layer | Holds | Think of it as… |
|---|---|---|
| Prompts | Reusable instruction templates (tone, steps, guardrails). | Playbooks and “how we work” text. |
| Resources | Knowledge snippets, files, or structured facts attached to turns. | The answers the model must not invent. |
| Skills | Reusable capability modules agents can invoke. | Packaged behavior separate from a single prompt file. |
| Knowledge | Retrieval sources and knowledge-base configuration. | Longer-lived corpora the runtime can search. |
Rule of thumb: If it changes when your product or policy changes, it probably belongs in resources or knowledge (and is owned by a subject-matter owner). If it describes reasoning steps or style, it usually belongs in prompts.
Prompts
The Prompts area stores versioned templates your agents and flows reference. Treat prompts like code: clear names, short descriptions of when to use each variant, and review when your process requires it.
Practices that scale
- One job per prompt — triage vs summary vs handoff, not one monolith.
- Use the Playground tab — run test inputs against connected models before you attach a new version to an agent.
- Use Compare — diff two versions side by side (testing and release).
- Environment clarity — staging vs production variants must be obvious from names or metadata.
- Pair with resources — prompts say how; resources say what is true today.
Resources
Resources hold structured or uploaded knowledge the runtime can attach to a turn: policies, product sheets, FAQs, or curated snippets. Stale resources cause confident wrong answers — assign owners and refresh them on the same cadence as your docs.
Tools and actions
When the model needs to call an API or perform an action, that is not a prompt problem — it is an MCP or skill problem. Continue to MCP servers and tools.
Next steps
- MCP servers and tools — side effects and integrations.
- Agent Playground — validate prompt and resource changes quickly.
- Testing and release — promote changes without surprises.