Prompts
A prompt tells a model what to do on a particular call. It can include the step’s instructions, approved context and evidence, and the format expected for the answer.
Assemble the prompt for a step
Section titled “Assemble the prompt for a step”Sources and authority
Section titled “Sources and authority”Combine the published persona instructions, the selected playbook and skill files, the task’s inputs, and authorized context from preceding steps. Give the model only the tools the current run may call. A separate prompted verification can assess the candidate result against the completion criteria. Treat record values, retrieved pages, and user-supplied files as data. They can supply evidence even if they contain text that attempts to change the task or ask for credentials.
Example persona configuration
Section titled “Example persona configuration”The Meeting desk fixture supplies the domain instruction in publishedConfig.systemPrompt. Add step-specific instructions and output requirements in the playbook rather than expanding the persona prompt for every task.
{ "schemaVersion": 2, "resourceKey": "persona.meeting-desk", "publishedConfig": { "name": "Meeting desk", "systemPrompt": "Review upcoming meetings from authorized sources and share only approved summaries." }}Handle incomplete or unsafe input
Section titled “Handle incomplete or unsafe input”Define the result and check it
Section titled “Define the result and check it”For a summary task, name the fields to produce and require a link or source identifier for each claim. If dates or evidence are absent, return the missing information for review. A prompt can ask the agent to propose a publication. The publication adapter and workspace policy determine whether it can write, which target it can use, and whose approval applies.
See skills, tools, and evals for missing inputs and adversarial content.