Bili Content Analysis

Choose an available interface

• Use the user's requested interface. Otherwise use connected BiliStalker MCP tools, or the CLI when terminal access and the project are available.
• For CLI invocation, read references/cli.md. Discover tool names with bili-stalker-mcp tools and inspect arguments with bili-stalker-mcp tools TOOL; the CLI shares MCP tool names, schemas, and results.
• Use credentials configured for the chosen runtime. An MCP client's environment settings do not automatically apply to a separate terminal.

Run workflow

1. Clarify target and scope

• Extract target identifier from user input (uid/username, bvid, article_id, keyword). Preserve a video's full URL including its p parameter, or pass the intended chapter with subtitle_page/subtitle_cid.
• Video inputs accept a BVID, AV number, full video URL, or https://b23.tv/... short link; the tool resolves them. If short_link_resolution_failed is returned, use an available full video URL or BVID, or request one from the user. For invalid_video_identifier, correct the input before retrying.
• Resolve usernames with search_users and reuse the selected numeric UID. Ask when candidates cannot be distinguished from the user's context.
• Keep response language aligned with the user unless explicitly requested otherwise.
• Ask for missing identifiers only when tools cannot resolve them safely.

1. Collect minimum sufficient evidence

• For a broad user overview, start with get_user_snapshot; set unneeded section limits to 0 and check errors for incomplete sections.
• For a specific video or article, fetch its detail directly. Video transcripts require fetch_subtitles=true; both smart and full default to P1 unless a chapter is selected. Use smart for one best language track or full for all language tracks within the selected scope, with a suitable subtitle_max_chars. For the whole video, set subtitle_scope="all" and use a BVID or URL without p.
• Use get_user_videos, search_user_videos, get_user_dynamics, or get_user_articles for targeted discovery and deeper pagination. Prefer lightweight lists and fetch heavy detail only for selected items.
• For video clues in recent dynamics, use get_user_dynamics with dynamic_type="VIDEO". Use video list/search tools for upload history, totals, or keyword search; dynamics provide only partial coverage. Video dynamics carry top-level type="VIDEO" and video; origin.video belongs to a repost and may be another creator's work.
• For discussion analysis, use get_content_comments and fetch a selected thread with get_content_comment_replies. Note-style comments may be previews; retrieve note.cvid with get_article_content when full text is needed.

1. Reconstruct source before interpreting

• Preserve source structure by timeline, chapter, or original logic order.
• Correct obvious transcription noise only when context provides strong evidence.
• Distinguish speakers when style, viewpoint, or surrounding context supports separation.

1. Build the analysis

• Explain facts, logic chain, assumptions, evidence, themes, and recent shifts.
• Describe recent patterns from the observed sample; infer a change over time only when earlier material provides a comparable baseline.
• Mark inferred conclusions explicitly when they are not directly stated by the source.
• Preserve original wording and tone where it carries analytical value.

1. Keep output useful for downstream tasks

• Keep key anchors in output: uid, bvid, article_id, publish times, and key source snippets.
• Avoid fragmented one-line bullet dumps; keep coherent narrative blocks with clear headings.

1. Handle failures safely

• If required data is missing or retrieval fails, state exact blockers and stop speculation.
• Treat partial snapshots and unavailable subtitles as incomplete evidence. Check subtitles.scope for the selected chapter and returned text pages, alongside errors and truncated; a returned page can contain partial text. Correct invalid_subtitle_selection before retrying. Respect a returned retry_after instead of immediately repeating blocked queries.
• A 412 can also reflect the shared circuit being open: switching tools does not bypass its cooldown. Use successful snapshot sections or already retrieved dynamics where sufficient, and state the missing coverage. After the cooldown, dynamics may provide recent video clues if the video list remains unavailable; keep their source and limits explicit.
• Provide concrete next actions (for example: request missing id, retry with cursor/page, fetch detail for candidate items).

Output contract

• Start with the final analysis directly when context is sufficient.
• Skip method narration unless the user asks for it.
• Match depth to intent: brief asks get concise conclusions plus key evidence; deep asks get full reconstruction and multi-dimensional interpretation.

Load detailed style rules on demand

• Read references/analysis-style.md when detailed style constraints are required.
• Follow system and developer instructions first if any conflict appears.