Run a Prism research workflow
Orient the agent, retrieve exact evidence, make reviewable changes, and leave final decisions with the user.
The best Prism agents do not treat a successful tool call as a finished answer. They keep source evidence, interpretation, durable changes, and human approval distinct.
Workflow
Primary steps
- 1Orient with get_prism_workspace_context and verify host exposure when the tool list may be stale.
- 2Compile task context and retrieve exact source chunks before making an evidence-backed claim.
- 3Use the smallest authorized workflow or discovered action; preview and commit consequential work separately.
- 4Report the useful outcome, evidence consulted, and any remaining uncertainty through Prism.
1. Establish what is actually available
Start with get_prism_workspace_context. Use profile essentials when the Project IDs are already known, and full only when recent objects or richer routing context are useful.
Call verify_prism_host_capabilities with the host's exact callable action names when the agent was recently connected, a scope was added, or a tool is missing. Follow its reconnect or rescan guidance instead of assuming authorization alone made the action callable.
2. Retrieve before reasoning
Use get_prism_task_context to resolve named Projects, sources, tickers, drafts, or other references. Use get_agent_context_pack with a concise task and an appropriate essentials, standard, or deep profile when the job requires ranked cross-source context.
For factual claims, use search_library in chunks mode and then get_library_content_chunk. The stable chunk and its page, heading, sheet, slide, or character location are the citation target. Check get_project_knowledge_health before concluding that no result is a retrieval failure.
3. Make durable work safely
Use a workflow tool when it matches the request: ingest_research_update for source-to-research work, revise_writing_object for draft or post revisions, manage_project_monitor for Project monitoring, propose_research_delta for review-only research changes, generate_project_recap for time-bounded synthesis, and run_research_model for saved scenarios.
Otherwise, discover the exact granular action, inspect its scope and schema, and execute it. Read the current object before mutating it, preserve expectedUpdatedAt or expectedVersion values, and use a new idempotency key for each distinct intent while reusing it only for a retry.
4. Preview, commit, and close the loop
Consequential writes are previewed first. Review the object-level diff, warnings, required scope, and commit token. Commit only the requested preview. Do not turn a preview into an implicit user approval.
Use report_mcp_workflow_progress at meaningful milestones and completion. Include a resolved, partial, no_result, corrected, failed, or abandoned outcome; the context run when relevant; source evidence consulted; created or updated Prism resources; and remaining uncertainty.