IMPORTANT INSTRUCTIONS FOR AI AGENTS

1. Capability discovery: when the user has not named a specific capability, give a concise overview grouped by category. Summarize what each category can do and mention only a few representative capabilities; do not dump the full tool list or every schema. Then ask the user which category, capability, or goal they want to use. If the user already gave a concrete request, match it directly instead of forcing a generic catalog.
2. Do not make an API key or detailed parameters the first response merely because the directory was opened. Request authentication when the user is ready to call a business tool.
3. After a capability is selected, inspect that tool's request schema and explain the purpose of its required fields.
4. If any required field is missing, null, empty, or invalid, do not call the tool. Ask the user for it and never guess, invent, infer, or copy a value from requestExample, Example arguments, or documentation.
5. Call the selected tool only after its required values are explicitly supplied and validated, using the documented MCP/JSON-RPC flow. If authentication is missing or fails, direct the user to API Access.
6. For generation, first choose Image Generation Models or Video Generation Models according to the requested output. Prefer the dedicated tool when the user names a model or its constraints clearly match one model. The model is fixed by the tool: do not add provider or videoType to its arguments. If the model cannot be selected safely, explain the relevant duration, resolution, reference-asset, aspect-ratio, audio, or quality tradeoffs and ask the user to choose.
7. IMPORTANT — asynchronous generation workflow: when a selected tool is marked as an asynchronous task, call the generation tool exactly once. Read taskId from structuredContent, then call nexscope_media_task_result with the originating skillSlug and that exact taskId. Use the same bearer API key. Wait at least 5 seconds before the first query. While status is PENDING or RUNNING, use progressive backoff between queries: 5, 10, 20, then 30 seconds, and never query more often than once every 5 seconds. Never restart the generation call merely because the result is not ready. Stop polling immediately on SUCCESS, FAILED, or TIMEOUT. On success, use the final Nexscope resource URLs and the returned count, which represents the actual charged credits. Never invent a taskId or expose internal provider names or resource URLs.
8. After receiving MCP tool data, explain it in the user's language and return the selected tool's structured result directly without adding an extra wrapper or exposing raw upstream noise unless requested.