菜单
Diagnose repeated or unclear LLM Wiki failures by tracing symptoms to schema, source quality, workflow, structure, retrieval, or tool configuration. Use when lint does not explain the problem, Agent behavior drifts, or the same failure survives multiple rule changes.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
Turn existing source pages into reusable entity, concept, topic, comparison, or synthesis pages while preserving evidence and avoiding duplicate nodes. Use when the vault has ingested sources but needs cross-source structure, comparisons, or a maintained knowledge graph.
Inspect an LLM Wiki vault in read-only mode, summarize its current state, and route the user to the smallest appropriate wiki workflow. Use when entering a vault, asking what to do next, requesting a status overview, or when the correct wiki skill is unclear.
Draft an external-facing report, memo, proposal, or presentation from an LLM Wiki with traceable evidence, uncertainty labels, and a pre-submission review. Use when the user asks to turn vault knowledge into material for other people or formal submission.
Ingest one or a small batch of raw documents into an existing LLM Wiki by creating source pages, updating related pages, index, and append-only log with traceable evidence. Use when the user adds material, asks to process a PDF/article/note, or wants a source incorporated into the wiki.
Initialize a new Obsidian LLM Wiki vault with a concise CLAUDE.md, immutable raw layer, wiki directories, index, log, and one selected scenario. Use for a new vault, missing schema, or a request to set up the Karpathy LLM Wiki structure.
Run a read-only health check on an LLM Wiki for broken links, orphan pages, duplicate concepts, missing evidence, stale synthesis, index drift, and schema conflicts. Use after ingestion, on a maintenance cadence, or before trusting a major synthesis.
基于 SOC 职业分类
| name | wiki-diagnose |
| description | Diagnose repeated or unclear LLM Wiki failures by tracing symptoms to schema, source quality, workflow, structure, retrieval, or tool configuration. Use when lint does not explain the problem, Agent behavior drifts, or the same failure survives multiple rule changes. |
Find the root cause before proposing more rules. Remain read-only unless the user separately approves a repair.
Provide:
Avoid recommending a full rebuild until a smaller experiment has failed. Do not add several rules at once because that hides which change worked.
