wiki-ingest Skill

Ingest content into your local LLM-WikiMind knowledge base — a production implementation of Karpathy's LLM Wiki pattern.

Prerequisites

1. LLM-WikiMind installed — see setup guide
2. wiki-kb MCP server registered in CatDesk/OpenClaw
3. qmd installed: pip3 install qmd

Trigger phrases

• "Add this to my knowledge base"
• "Ingest this article"
• "Save to wiki" / "Write to knowledge base"
• "Store this note"
• wiki_ingest
• 加到知识库 / 写入知识库 / 保存到 wiki / 记录到知识库 / 把这个存起来

Workflow

Step 1: Read global schema

cat "$WIKIMIND_ROOT/CLAUDE.md"

Step 2: Classify the content

Step 3: Identify the domain

• Use an existing domain (check CLAUDE.md for the list)
• Or create a new one: make a directory + DOMAIN.md with keywords frontmatter

Step 4: Write the page

Preferred — use the MCP tool (if wiki-kb server is loaded):

wiki_ingest_note(
  title="Page Title",
  content="# Page Title\n\nContent here...",
  domain="my-domain",
  page_type="concept",
  source="https://example.com",
  summary="One-line summary",
  tags=["tag1", "tag2"]
)

Fallback — write the file directly:

---
title: "Page Title"
type: concept
domain: my-domain
source: "https://example.com"
summary: "One-line summary (<=150 chars)"
tags: ["tag1"]
related: []
created: YYYY-MM-DD
updated: YYYY-MM-DD
confidence: medium
---

# Page Title

Content here.

Save to: $WIKIMIND_ROOT//concepts/page-title.md

Step 5: Update the search index

cd "$WIKIMIND_ROOT" && qmd update

Step 6: Append to log

echo "## [$(date +%Y-%m-%d)] ingest | Page Title | domain" >> "$WIKIMIND_ROOT/<domain>/log.md"

Quality Standards

concept pages must include: definition, motivation, code example, common pitfalls, related links

entity pages must include: description, key properties with types, key methods with signatures, usage example, gotchas

comparison pages must include: comparison table, when to use each, migration notes

source-summary pages must include: original URL, key takeaways (3-7 points), what this adds vs existing knowledge

Notes

• Never write into refs/ — that's for bulk-imported raw docs
• confidence: high for official docs, medium for your summaries, low for speculation
• After writing, tell the user the file path and a good search query to find it