liyading1818/obsidian-llm-wiki

Wiki Skill

This skill turns the agent into a disciplined maintainer of the user's personal knowledge base at <VAULT_ROOT> (the repository root).

Step 0 — Always read AGENTS.md first

Before doing anything, read AGENTS.md at the vault root. It defines the directory layout, page conventions, frontmatter schema, log format, and detailed workflows. This SKILL.md is only a dispatcher; AGENTS.md is the source of truth.

Dispatch: what is the user asking?

Pick exactly one of these intents from the user's message:

User intent	Workflow
Asking a question, wants analysis / synthesis	Query (below)
Dropped a new file or said "ingest this"	Ingest (see AGENTS.md §4.1)
Said "lint" / "health check" / "find problems"	Lint (below)
Wants a comparison / slide deck / chart / report	Generate artifact (below)
Just exploring or unsure	Run wiki stats then ask what they want

If unclear, ask one short clarifying question before proceeding.

Query workflow

1. Run wiki search <keyword> in the repo root to find candidate pages. Use multiple keyword variations if needed.

2. Read wiki/index.md to see the catalog.

3. Read the candidate pages found in step 1.

4. If still missing context, fall back to reading the relevant files in raw/ or clippings/.

5. Synthesize an answer with [[wikilink]] citations to wiki pages and [文章](../raw/xxx.md) citations to raw sources.

6. If the answer has lasting value (a comparison, a derived insight, a non-trivial synthesis), write it as a markdown file to output/<YYYY-MM-DD>-<short-slug>.md with frontmatter:

   ---
   type: artifact
   kind: answer | comparison | analysis | report | slides
   created: <date>
   question: "<the user's question, verbatim>"
   sources_used: [wiki/concepts/X.md, wiki/sources/Y.md, ...]
   ---
   

   Then ask the user: "要不要把这份输出归档为 wiki/questions/<slug>.md？"（If yes, copy + update index.md & log.md.）

7. If the answer is trivial (a fact lookup), just answer in chat — no file needed.

Lint workflow

1. Run wiki lint — this writes lint/<today>.md automatically.
2. Read that file and summarize for the user (which issues are real, which are false positives like literal [[examples]] in docs).
3. Propose a prioritized fix plan. Don't auto-fix — let the user pick.

Generate artifact workflow

Same as Query, but the deliverable form is explicit (table / Marp slides / Python chart / canvas). Always save to output/. For slide decks, use Marp markdown format. For charts, write a Python script to output/scripts/ and run it to produce the image alongside.

Output directory convention

output/
├── 2026-06-26-rag-vs-llm-wiki.md          # answer / comparison artifact
├── 2026-06-26-karpathy-summary.slides.md  # Marp slide deck
├── scripts/                                # helper scripts for chart generation
└── assets/                                 # rendered images, exports

The output/ directory is LLM-owned, append-only (don't delete prior artifacts without asking). Each file must have the frontmatter shown above so the user can later promote artifacts into the wiki.

CLI commands available (run from <VAULT_ROOT>)

Only four commands, deliberately minimal:

wiki stats              # overview: page counts by type/status
wiki log -n 10          # last N entries from wiki/log.md
wiki search <regex>     # full-text search (--scope wiki|raw|clippings|all)
wiki lint               # write lint/<today>.md with orphans/broken/stubs

Everything else (creating pages, editing frontmatter, updating index.md, appending to log.md) is done by directly reading and writing markdown files. Don't ask for a CLI to do file edits — that's your job.

Critical rules (from AGENTS.md)

• Never modify files in raw/ or clippings/.
• Every ingest must update wiki/index.md and prepend an entry to wiki/log.md with format ## [YYYY-MM-DD] ingest | <title>.
• Cross-reference with [[double-brackets]] between wiki pages.
• Pages need YAML frontmatter (type, tags, created, updated, sources, status).
• When uncertain, ask the user. Don't do silent large refactors.

Reporting style

After completing a workflow, give the user a short summary:

• What you did (1–3 bullets)
• What artifacts were created or modified (relative paths)
• Any unresolved questions or new gaps you noticed

Keep it tight. The user is browsing Obsidian in parallel and can see file changes.