| name | llm-wiki-starter |
| description | TRIGGER PHRASES — match ANY of these (case-insensitive):
· 创建 llm-wiki 知识库 / 创建知识库 / 创建 wiki / 创建 wiki 知识库 / 创建本地知识库
· 搭建 llm-wiki / 搭建 wiki / 搭建知识库 / 搭建本地知识库 / 搭一个 wiki
· create llm wiki / scaffold llm wiki / set up llm wiki / build a local wiki / scaffold a knowledge base / set up obsidian wiki / start a new wiki vault
· /llm-wiki-starter (slash command)
Scaffolds a NEW LOCAL Obsidian-based LLM Wiki on the user's machine, following Andrej Karpathy's pattern. End-to-end: installs Node/Git/jq, kepano/obsidian-skills, axtonliu/visual-skills, Obsidian app, 17 plugins + Minimal theme; creates a fresh vault from the official template; initializes git. Each invocation produces a NEW vault — re-runs prompt for a fresh name.
DO NOT match for: 飞书/Lark wiki (lark-wiki handles that), online/hosted wikis (Notion, Confluence, GitHub wiki), ingesting content INTO an existing wiki (wiki-ingest), or generic Obsidian-vault editing (obsidian-markdown / obsidian-cli).
|
llm-wiki-starter
Guide the user through scaffolding an LLM Wiki knowledge base end-to-end: detect what's already installed, install what's missing across OS/package-manager combinations, create the vault from the official template, and configure Obsidian.
The user is already running an AI CLI to invoke this Skill — do NOT prompt them to install another AI agent.
When to use
- Natural language: user says 创建 llm-wiki 知识库, 创建知识库, create llm wiki, scaffold llm wiki, set up knowledge base.
- Explicit command:
/llm-wiki-starter (optionally with --bash, --name <wiki-name>, --lang <en|zh>, --dir <path>, --only-tools, --only-wiki, --only-obsidian). See Parameter reference below for full semantics.
If explicit command is used with parameters, skip the matching interactive prompts.
Global principles (apply to every stage)
- Detect before install. Each step: check if target exists → print ✓ skip / ↓ install / ⚠ fail. Never reinstall.
- Failures do not block. On single-item failure, record it to a "manual install" list and continue. Print the full list at the end.
- Respect OS and shell. Detect with
uname -s (macOS / Linux / Windows-via-Git-Bash). If running under native Windows cmd/PowerShell (no uname), use winget/choco/scoop directly — do not assume bash.
- Use the host CLI's native interaction. If you are Claude Code, use AskUserQuestion. If Codex or another CLI, use its native prompt mechanism. Do not hardcode a UI shape.
- Idempotent for tools, NOT for wiki creation. Re-running detects already-installed tools and skips them. But Stage 4 (create wiki) is the workflow's purpose: every invocation MUST prompt for a new wiki name / lang / dir and produce a NEW vault. The existence of a previously-created wiki (e.g.
~/Documents/CODE/my-wiki) is NOT grounds to skip Stage 4 — ask the user for a new name and create a fresh vault next to it.
- Never prompt for login / API keys / tokens / extra AI CLIs. The user is already running an AI CLI — leave their AI tooling alone.
- Windows line endings. When writing files the user will commit, use LF (Unix) line endings.
- Act, don't narrate. Don't pre-announce what you're about to do. Don't recap which Stage you're in. Don't say "需要确认 X" / "I'll now ask you about Y" / "进入 Stage N" / "稍后会提醒"; just take the action — ask the question, run the command, write the file. The user only needs ONE concise line per outcome (
✓ <name>, ⚠ <name>: <reason>), not a play-by-play.
- Report each item individually, not in aggregate. "✓ kepano/obsidian-skills installed" + "✓ axtonliu/visual-skills installed" — not "Both Agent Skills are installed". Lists of items get one line each.
- Each warning appears ONCE. Don't pre-warn at detection time and then re-warn at finalize. If something needs the user's attention, surface it in the final summary only — not midway through the run.
Workflow (6 stages)
Execute stages in order. Read the matching reference file before each stage. Do not skip reference files — they contain exact commands.
| # | Stage | Reference |
|---|
| 0 | Entry alignment (OS detect, plan, confirm) | SKILL.md (this file, section below) |
| 1 | Base tools (Node / Git / jq / curl) | references/02-install-base.md |
| 2 | Agent Skills (kepano + axtonliu) | references/03-install-skills.md |
| 3 | Obsidian app + Web Clipper | references/04-install-obsidian.md |
| 4 | Create wiki from template | references/05-create-wiki.md |
| 5 | Obsidian plugins + Minimal theme | references/06-install-plugins.md |
| 6 | Finalize (git, summary) | references/07-finalize.md |
Detection rules shared across stages: see references/01-detect-tools.md.
Stage 0: Entry alignment
-
Run uname -s to detect OS. Record as OS = macos / linux / windows.
-
Parameter hint (only when triggered by bare /llm-wiki-starter with NO parameters): print ONE concise line listing available flags before anything else, so the user discovers them. List --bash first since power users invoke it most often:
Parameters (optional): --bash --name <wiki-name> --lang <en|zh> --dir <path> --only-tools --only-wiki --only-obsidian
No params given — continuing interactively.
Skip this hint if: (a) the user passed any parameter, OR (b) triggered via natural language (not the slash command).
-
Parse any CLI parameters passed with /llm-wiki-starter. Remember them for later stages. If an unknown flag appears, print ONE line: Unknown parameter: <flag>. Valid: --bash, --name, --lang, --dir, --only-tools, --only-wiki, --only-obsidian. Ignoring. and continue. The three --only-* flags are mutually exclusive — if more than one is passed, keep the last and warn Conflicting --only-* flags; using <last>.
-
If --bash was passed: short-circuit to bash mode. Do NOT read any other reference files, do NOT run the 6-stage SOP. Instead, run a single bash command that pipes the upstream install.sh into bash, forwarding the other parameters:
curl -fsSL https://raw.githubusercontent.com/eleven-net-cn/llm-wiki-starter/main/install.sh \
| bash -s -- --yes <forwarded params>
Forward only parameters install.sh understands: --name, --lang, --dir, --only-tools, --only-wiki, --only-obsidian. Always append --yes so install.sh runs non-interactively.
Example: /llm-wiki-starter --bash --name foo --lang zh → runs curl ... | bash -s -- --yes --name foo --lang zh.
After install.sh exits, relay its exit status (success / failure) in ONE line. Do NOT layer a Skill-style summary on top — install.sh prints its own summary already.
-
Go directly to Stage 1. Do NOT ask "Proceed?" or any other confirmation — the user's request to create a wiki IS the confirmation. Global principle #8 (Act, don't narrate) applies to trigger paths too.
Reminder: even if the user has run this Skill before and has an existing wiki on disk, the goal of this run is to create a NEW wiki. Stage 4 will prompt for a fresh name/dir. Do not interpret a prior wiki as "the work is already done."
Parameter reference
Names mirror install.sh so the two install paths stay aligned. Listed in suggested-display order (most commonly used first).
| Flag | Default | Behavior |
|---|
--bash | (off) | Bypass the 6-stage SOP. Execute upstream install.sh in one shell call (non-interactive, --yes) and forward the other flags. Minimizes AI token cost. The AI reads only SKILL.md Stage 0 step 4, no other reference files. |
--name <wiki-name> | (prompt) | Wiki directory name. Collision → re-prompt. |
--lang <en|zh> | en (or prompt) | Language overlay for template. |
--dir <path> | $(pwd) (or prompt) | Parent directory; wiki created at <dir>/<name>. |
--only-tools | (off) | Run Stages 1-3 only (base tools / Agent Skills / Obsidian app). Skip Stages 4-6 — no wiki created. |
--only-wiki | (off) | Skip Stages 1-3 (assumes tools already installed). Run Stage 4 (create wiki) + Stage 5 (plugins/theme) + Stage 6 (finalize). |
--only-obsidian | (off) | Skip Stages 1-4. Run Stage 5 only (install plugins + theme into an existing vault) + a brief Stage 6 summary. Requires --dir pointing at a vault containing CLAUDE.md. |
Then proceed to stage 1 (or skip ahead per the --only-* flag in effect).
Canvas constraint (already encoded in template)
The template's AGENTS.md instructs future AI sessions inside the new vault to use obsidian-canvas-creator (axtonliu/visual-skills) for .canvas creation, not json-canvas (which kepano/obsidian-skills bundles). You don't need to repeat this in the user-facing finalize summary — that line is for AI coordination, and AGENTS.md already carries it into the vault.
