Prompts
Prompt library
A prompt in Dezifi is a named, versioned, reusable instruction set. Manage it once; agents reference it by version. When the prompt evolves, agents pick up the change through a controlled lifecycle — not a copy-paste.
What you'll learn
- How prompts differ from an agent's inline system prompt
- The status lifecycle of a prompt version
- How prompts attach to agents through Skills
- When to extract content into a prompt versus keep it inline
Prompt vs. agent system prompt
Every agent has an inline system prompt — short, agent-specific, defines its role. A library prompt is different: it lives in the prompt library, is versioned with its own lifecycle, and can be attached to many agents through Skills. Use the inline system prompt for what makes this agent unique. Use a library prompt for instructions you want to reuse, govern, and version centrally — style guides, classification rubrics, refusal language, brand voice.
The version lifecycle
- 1
Draft
Editable. Not attachable to a production agent. Where you author and iterate. - 2
Active
Production-ready. Attachable to agents. The default version when an agent references the prompt without pinning. - 3
Deprecated
Still works for existing attachments but cannot be attached to new agents. Signal to consumers to migrate to a newer version. - 4
Archived
Read-only history. Cannot be attached or activated. Kept for audit and rollback reference.
How agents consume prompts
Agents pull library prompts in through the Skill Selection step of the agent builder. Skills wrap a prompt plus optional tool bindings and retrieval hints, so the agent gets a coherent unit of behavior rather than raw text. The agent pins a specific version or follows the Active version — pick the model that matches how often you want changes to roll forward automatically.
Frequently asked questions
- Should every system prompt live in the library?
- No. Keep agent-unique role definitions inline. Extract to the library only when the same instruction shows up across multiple agents, or when you want to version and govern it independently of any one agent.
- Can a prompt depend on variables?
- Yes. Prompt content supports template variables that the agent or workflow fills at runtime — caller name, current date, tenant context. Variables are declared at the top of the prompt so attachments know what to supply.
- Who can edit a prompt?
- Authors and admins. Editing is only possible on Draft versions — Active, Deprecated, and Archived versions are immutable, which keeps the audit trail clean.
- How do I see which agents use a prompt?
- Open the prompt detail page. The Used by tab lists every agent and workflow that attaches the prompt, pinned to which version. Use it to plan migrations before deprecating.