Explore Prompts

Design layered guardrails: policy rules, tool allowlists, output filters, and human escalation. Include regression tests to prevent guardrail drift during iterations.

Design end-to-end tracing: prompt versions, retrieval results, tool calls, retries, and final answers. Include a log schema and query examples for incident investigation.

Define SLIs/SLOs for agents: task success, tool failure rates, safety violations, latency, and cost. Provide a dashboard layout and alert thresholds.

Create a decomposition framework for multi-hour tasks: milestones, checkpoints, intermediate artifacts, and backtracking rules. Include a “progress ledger” to prevent losing the plot.

Design a verifier model workflow: primary proposes, verifier checks constraints and factuality, policy gate decides. Include disagreement handling and confidence thresholds.

Design a critic pattern: structured critique prompts, bounded iterations, and a stop rule. Include how to measure improvement vs added latency/cost.

Create an A2A negotiation mechanism: task contracts, acceptance criteria, SLAs, and cost budgets. Include how agents refuse tasks they cannot verify or safely perform.

Design MCP tool schemas: typed inputs/outputs, explicit error codes, idempotency rules, and logging fields. Provide examples of good vs bad schema design for agent reliability.

Create a sandbox model for tool execution: network isolation, file access boundaries, rate limits, and safe parameter validation. Include a threat model and tests for containment failures.

Design a RAG pipeline for agents: retrieval, citation, cross-checking, and contradiction handling. Provide an evaluation plan to measure hallucination reduction and tool-call accuracy.

Create a memory architecture: scratchpad, episodic memory, semantic memory, and “facts of record.” Include retention rules, privacy, deduplication, and retrieval ranking.

Design a planner–executor split: planner produces a verifiable plan, executor performs bounded actions, verifier checks constraints. Include gate criteria, rollback triggers, and audit logging.