Prompt Cards

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.

Design an agent-to-agent (A2A) coordination protocol: message types, handoffs, escalation rules, and consensus patterns. Include how to prevent loops, duplication, and hallucinated delegation.

Create an MCP-first design: define servers, tools, schemas, and a permission model. Provide naming conventions, versioning, and a “deny-by-default” policy for tool exposure.

Design a high-capability agent architecture (marketed as “IQ150-level”) that stays reliable: modular planner/executor, tool sandboxing, and human-in-the-loop gates. Include a component diagram, data flows, and failure-mode mitigations.

Create a 90-day implementation plan for a safe recursive improvement system: milestones, deliverables, tests, monitoring, governance, and a pilot rollout. Include a ‘pause and assess’ week.

Design a weekly meeting agenda focused on safety and reliability: review dashboards, incidents, near-misses, red-team findings, and upcoming risky changes. Include decision outputs and owners.

Create a compliance workflow: policy mapping, data handling rules, documentation, audits, and change impacts. Keep it general (not legal advice) and focused on operational safety.

Create a delegation pattern where the AI plans but humans execute high-risk steps. Include how to structure instructions, confirmations, and evidence requirements.

Design methods to evaluate long-horizon tasks: checkpoints, intermediate artifacts, verifier models, and human spot checks. Include metrics that detect slow drift or hidden degradation.

Create a security review plan for the AI toolchain: dependencies, container images, permissions, SBOM, and update policy. Include a checklist and a cadence for reviews.

Define hard caps on recursion: max iterations, time limits, budget limits, and scope constraints. Provide rules for when to reset context and when to escalate to humans.