wiki-ingest

Process a new source into an existing Karpathy-style LLM wiki. The pattern and the ingest example this skill follows are described in https://gist.github.com/karpathy/442a6bf555914893e9891c11519de94f.

Karpathy's ingest example, verbatim:

the LLM reads the source, discusses key takeaways with you, writes a summary page in the wiki, updates the index, updates relevant entity and concept pages across the wiki, and appends an entry to the log.

A single ingest may touch 10–15 pages. Karpathy prefers one source at a time with the user in the loop; batch mode is also possible (see the end).

Prerequisites

• A wiki exists. If not, run wiki-init first.
• The llm-wiki-mcp server is wired into this session. If

wiki_inventory is not callable, tell the user to add the server to their MCP client config and restart.

Pre-flight

Read CLAUDE.md. This is the authoritative schema — page categories, frontmatter fields, link conventions, and operation vocabulary. The user may have evolved it since wiki-init. Honor the current state, not the defaults.

Flow

Follow Karpathy's six actions in order.

1. Read the source

Use whatever the host agent provides — WebFetch for URLs, Read for local files, or treat pasted text as-is. The llm-wiki-mcp server does not fetch sources; extraction is the host's job. Archiving the raw file into raw/ is the user's job, not this skill's.

2. Discuss key takeaways with the user

Surface the handful of entities, concepts, and claims that matter. Ask the user what to emphasize.

3. Write a summary page in the wiki

Pick a slug (lowercase, dashes, at least two characters). Build body to match the schema in CLAUDE.md, including any frontmatter block. The server does not validate frontmatter shape — you are responsible for consistency.

wiki_write_page(slug=..., body=..., etag=None)

etag=None asserts the page does not exist. If the server returns a conflict error, the slug is taken — pick a different one.

4. Update the index

index.md sits outside pages/, so the MCP tools do not touch it. Read it with the host Read tool, add a one-line entry under the correct category section, write it back with Write.

5. Update relevant entity and concept pages across the wiki

Extract the set of key terms from the source — entity names, concept names, notable claims. Aim for five to twenty terms.

wiki_inventory(scan_for=<terms>)

The response's mentions field lists each existing page that mentions any of the terms and the exact line. The pages field tells you which entities and concepts already have their own page versus which are only mentioned in passing and deserve one now. Note that wiki_inventory does not return page bodies — only metadata and mention lines. Use wiki_read when you need the full body of a page to update.

Use this as the ground truth for Karpathy's "relevant entity and concept pages":

• Existing page, new information. wiki_read(slug=...) for the

current body and etag. Integrate the new claim, source citation, backlink, or contradiction note. wiki_write_page(slug, body, etag=<previous etag>). On conflict, re-read and retry. Never discard existing content.

• New entity or concept deserving its own page. Create it with

wiki_write_page(..., etag=None), then add it to the index (repeat step 4 for the new slug).

Include [[slug]] backlinks to other wiki pages wherever natural. Link only to slugs that appear in wiki_inventory.pages.

6. Append a log entry

wiki_log_append( operation="ingest", title=<source title>, extra_lines=[<one or two short lines>], )

The server formats the entry as ## [YYYY-MM-DD] ingest | <title> and guarantees atomic append. Concurrent ingests will not clobber each other. Use operation="ingest" unless CLAUDE.md has defined different vocabulary.

Tool cheatsheet

The llm-wiki-mcp server deliberately owns only page CRUD, inventory, and log append. Index, raw/, and source fetching stay with the host agent.

Batch mode

If the user explicitly asks to batch-ingest a directory, loop the flow per source. Skip step 2's discussion unless the user wants it. Pause every five sources to report progress.

Reporting

When done, report: source title, pages created (slugs), pages updated (slug + one-line diff reason), and the log entry appended. Keep it terse — the user can read log.md or call wiki_inventory for details.