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。
直接命令不会经过审查 Prompt;运行前请先检查来源。
npx skills add https://github.com/D0miH/obsidian-second-brain-research-skills --skill audit-links命令会保持在同一行。复制前请横向滚动并检查完整内容。
想先保存到本地?可下载 SkillsMP 当前能够提供的文件。
正在显示 SKILL.md
基于 SOC 职业分类
| name | audit_links |
| description | Validate vault graph consistency and structural integrity |
Fast audit of vault structure to catch:
[[Attention Mechanism]])Vault constraints (non-negotiable):
Reading/ must link to ≥1 project in Projects/Projects/ must link to ≥1 MOC in MOCs/[[Transformer]], [[Attention Mechanism]])Use the Agent tool to spawn a subagent. Pass it the prompt from the SCAN SUBAGENT PROMPT section below verbatim. Wait for the subagent to return its structured findings before proceeding.
Parse the subagent's output and display violations grouped by severity:
Each violation: file name + issue + suggested fix. If subagent returned NO_VIOLATIONS, report clean audit.
Prioritize by severity (orphans > broken bidirectional > concept links). Offer fixes: "Should I remove this link?" / "Should I add reciprocal link?"
After completing the audit (regardless of whether fixes were applied):
date via Bash to get the current system time.claude/vault_maintanance.mdaudit_links: line to current date and time in format DD-MM-YYYY HH:MMpapers_since_audit_links: to 0projects_since_audit_links: to 0Pass this prompt verbatim to the Agent tool:
You are a vault structure auditor for an Obsidian research vault.
Vault root: {{VAULT_ROOT}}
STEP 1: Get audit scope
Read `.claude/vault_maintanance.md` and note the `audit_links:` timestamp.
Default: incremental scan — only files modified after that timestamp.
Convert the timestamp (DD-MM-YYYY HH:MM) to YYYYMMDDHHMI format for `touch -t`.
Run: touch -t YYYYMMDDHHMI /tmp/audit_ref && find "Reading/" "Projects/" "MOCs/" -newer /tmp/audit_ref
(Run this bash command from the vault root directory)
Only do a full scan if no audit_links: timestamp exists yet.
STEP 2: Scan for violations
Exclude: Inbox/, Daily Notes/
For Archive/: skip orphan checks, but include in bidirectional validation.
For papers in Reading/:
- Run: grep -rl "\[\[" "Reading/" to get files, then grep "\[\[" on each to extract links
- Flag: Papers with zero links pointing to files in Projects/ (orphan papers)
- Flag: Papers with any link pointing to a file in MOCs/ (forbidden)
- Flag: Concept links — single-concept terms with no paper-like specificity e.g. [[Attention]], [[Transformers]], [[Softmax]]
For projects in Projects/:
- Extract all [[wiki links]] from each project using grep
- Flag: Projects with zero links pointing to files in MOCs/ (orphan projects)
For bidirectional consistency:
- For each paper→project link: verify the project file contains a reciprocal link back to the paper
- Flag: Missing reciprocal links
STEP 3: Return findings in EXACTLY this format (nothing else):
ORPHAN_PAPERS:
- [[filename]] | no project links found | add link to relevant project
(or write "none" if no violations in this category)
ORPHAN_PROJECTS:
- [[filename]] | no MOC links found | add link to relevant MOC
(or "none")
BROKEN_BIDIR:
- [[paper]] → [[project]] | project missing reciprocal link | add [[paper]] to project's Relevant Papers
(or "none")
FORBIDDEN_CONCEPT_LINKS:
- [[filename]] | [[ConceptLink]] | remove link
(or "none")
FORBIDDEN_MOC_LINKS:
- [[paper]] | links to [[MOC Name]] directly | remove link
(or "none")
MALFORMED_LINKS:
- [[filename]] | malformed link text | fix
(or "none")
If zero violations across ALL categories, return only the single line: NO_VIOLATIONS
Scan work is fully isolated in a subagent — it does not pollute the main context window. Main session only handles user interaction and writes.
✅ Did I spawn the scan subagent and receive structured findings? ✅ Did I present findings grouped by violation type? ✅ Did I get user approval before fixing? ✅ Did I update vault_maintanance.md with current timestamp and reset counters?