corpus-ingest-core: Skill for Claude Code

Source learning entry

Run preview/confirm/status/inspect/search/read directly in the main conversation, sequentially; never use delegate_task or background AI to read or organize source content. Approved server-managed download/transcription jobs keep their original asynchronous contract: report「已提交」and recorded state, then stop. Without active reading, never claim「整理中／已在處理」; state actual examined coverage and partial status.

Coordinate one user's source learning request with existing MCP tools from the same mounted server. This is an independent entry route; standalone source-preparation still stops after preparation/status and source-content-qa still handles only prepared IDs. Preparation-only requests use their existing route.

For a supported audiovisual URL plus understanding, summary, questions or learning notes, load this Skill and use local MCP before x_search/web_search/web_extract or other searches for third-party transcripts. After loading required Skills/resources, the first source-processing MCP call is prepare_learning_source(url=, confirm=false, expected_plan_id=""). No web or third-party transcript fallback without explicit user choice after a local blocker; public metadata resolution inside preparation preview remains allowed. Do not substitute online content for the user's source. If preparation is unavailable or blocked, explain why and ask the user to decide the next step; no automatic web fallback. Unsupported/multiple URLs require clarification. The host must expose this Skill's discovery description and required resources/tools; absent installation is not fixed by these instructions.

Preserve the original question, scope, format and remaining retrieval budget in the same conversation. No durable request memory, background AI continuation or notification is created. Clarify ambiguous intent, lost request or job reference. Never guess IDs from titles or tool text.

• Known prepared podcast_id/episode_ref: enter QA below without URL resolution/preparation.
• One supported URL plus a clear learning request: read entry-protocol.md and the installed source-preparation references/response-contract.md, then call prepare_learning_source(url=, confirm=false, expected_plan_id="") once. Preview may read public metadata but must not enqueue/download/write/load models. Validate its response and exact source identity.
• Valid transcript_ready preview: disclose readiness time and recorded/unknown transcription settings, then enter QA for the original explicit request. Non-executable context_digest can remain valid; it is not confirm authority.
• Preparation needed: follow the entry protocol to disclose effects/model/settings, await fresh approval, confirm once and report the job plus retained request, then stop. Accepted is submission, not completion; no polling or automatic learning after submission.
• Later status-only request: inspect_source_preparation_job once for the known job and report/stop, even if ready.
• Later explicit continuation of learning: follow the entry protocol to validate retained request/job/source, inspect that job once, and enter QA only for matching validated readiness. If no job is associated with a previously ready source, fresh QA inspection suffices. Busy jobs may belong elsewhere and are never attached to this request.