Skip to content

Instantly share code, notes, and snippets.

@blackestwhite
Created April 10, 2026 22:50
Show Gist options
  • Select an option

  • Save blackestwhite/80969d33cc2afdb2aeaf7fbbef5635f7 to your computer and use it in GitHub Desktop.

Select an option

Save blackestwhite/80969d33cc2afdb2aeaf7fbbef5635f7 to your computer and use it in GitHub Desktop.
Andrej Karpathy's LLM Wiki inspired checklist

LLM Wiki Checklist

An actionable checklist for setting up and maintaining a personal knowledge base where an LLM incrementally builds and updates a persistent wiki from raw sources.

Goal

  • Decide the domain for this wiki.
  • Define the scope clearly enough that the LLM knows what belongs in the wiki and what does not.
  • Decide whether the wiki is primarily for personal use, research, books, team knowledge, or another use case.
  • Choose the LLM agent you want to collaborate with.
  • Decide where you will browse the wiki while the LLM edits it.

Core Model

  • Treat raw sources as immutable source-of-truth documents.
  • Treat the wiki as an LLM-maintained layer that sits between you and the raw sources.
  • Treat the schema file as the operating manual that tells the LLM how to maintain the wiki.
  • Commit to the idea that valuable synthesis should be stored in the wiki rather than lost in chat history.
  • Decide how much of the maintenance work the LLM should own versus how much you want to review manually.

Architecture

  • Create or choose a directory for raw sources.
  • Create or choose a directory for the wiki pages.
  • Create a schema file for your agent, such as AGENTS.md or CLAUDE.md.
  • Document the relationship between raw sources, wiki pages, and schema instructions.
  • Decide how the LLM should name, organize, and cross-link pages.

Wiki Structure

  • Define the main page types you want in the wiki.
  • Decide whether you want source summary pages.
  • Decide whether you want entity pages.
  • Decide whether you want concept pages.
  • Decide whether you want overview or synthesis pages.
  • Decide whether you want comparison pages for recurring questions.
  • Decide whether you want consistent frontmatter or metadata on pages.
  • Decide what categories should appear in index.md.

Schema Rules

  • Tell the LLM how to ingest a new source.
  • Tell the LLM how to answer questions from the wiki.
  • Tell the LLM how to log changes in log.md.
  • Tell the LLM how to update index.md.
  • Tell the LLM how to handle contradictions between old and new information.
  • Tell the LLM how to create and maintain cross-references.
  • Tell the LLM when to create a new page versus update an existing one.
  • Tell the LLM what citation style to use.
  • Tell the LLM what kinds of outputs can be filed back into the wiki.
  • Tell the LLM how to run periodic lint or health-check passes.

Ingest Workflow

  • Add a new source to the raw collection.
  • Ask the LLM to read the source.
  • Review the source's key takeaways with the LLM if you want a supervised workflow.
  • Create or update a source summary page.
  • Update relevant entity pages.
  • Update relevant concept pages.
  • Update synthesis or overview pages when the new source changes the big picture.
  • Flag contradictions, uncertainty, or superseded claims.
  • Update index.md.
  • Append an ingest entry to log.md.
  • Decide whether you want to ingest sources one at a time or in batches.

Query Workflow

  • Ask questions against the wiki instead of against only the raw sources.
  • Have the LLM consult index.md first to find likely relevant pages.
  • Have the LLM read the specific pages needed for the answer.
  • Require citations back to wiki pages and, when needed, raw sources.
  • Decide which query outputs should become new wiki pages.
  • Save valuable analyses, comparisons, or synthesized answers back into the wiki.
  • Update index.md and log.md when a query creates durable new knowledge.

Lint Workflow

  • Run periodic health checks on the wiki.
  • Check for contradictions between pages.
  • Check for stale claims that newer sources may have superseded.
  • Check for orphan pages with no meaningful links.
  • Check for important recurring concepts that still lack their own page.
  • Check for missing cross-references.
  • Check for obvious data gaps or unanswered questions.
  • Ask the LLM to suggest new sources or follow-up questions.
  • Append a lint entry to log.md.

Index And Log

  • Create index.md as a content-oriented catalog of the wiki.
  • Give each index entry a link and a one-line summary.
  • Optionally include metadata like dates, tags, or source counts in the index.
  • Organize the index by meaningful categories.
  • Create log.md as an append-only chronological record.
  • Use a consistent heading format for log entries so they are easy to parse.
  • Log ingests.
  • Log queries that created durable artifacts.
  • Log lint passes and maintenance work.

Optional Tooling

  • Start with plain markdown files and simple file-based workflows.
  • Decide whether the wiki is still small enough for index.md-driven navigation.
  • Add search tooling only when scale makes it necessary.
  • Evaluate tools like qmd if you want local markdown search with ranking.
  • Consider building small helper scripts only when a real workflow bottleneck appears.
  • Keep optional tooling modular so the wiki still works without it.

Obsidian Workflow

  • Decide whether Obsidian will be your primary interface for browsing the wiki.
  • Use Obsidian Web Clipper if you want fast article-to-markdown capture.
  • Decide whether to download images locally for clipped sources.
  • Configure a fixed attachment folder if you want reliable local image storage.
  • Decide whether the LLM should inspect images separately from the source markdown when needed.
  • Use graph view to inspect connectivity and identify hubs or orphans.
  • Consider Marp if you want the LLM to generate slide decks from wiki content.
  • Consider Dataview if you want structured metadata queries over the wiki.

Maintenance Habits

  • Keep raw sources unchanged after import.
  • Let the LLM do the repetitive bookkeeping work.
  • Review the wiki often enough to catch drift in structure or emphasis.
  • Refine the schema over time as you learn what works.
  • Keep the wiki in git if you want version history and easier experimentation.
  • Prefer persistent notes over ephemeral chat answers when the output has lasting value.

Human Responsibilities

  • Curate good sources.
  • Direct the analysis.
  • Ask strong questions.
  • Review important updates.
  • Adjust the schema when the workflow needs improvement.

LLM Responsibilities

  • Summarize new sources.
  • Update existing pages consistently.
  • Create new pages when needed.
  • Maintain cross-references.
  • Surface contradictions and uncertainty.
  • Keep indexes and logs current.
  • Help identify gaps, follow-up questions, and next sources to explore.

Customization Decisions

  • Decide how much structure you want versus how lightweight the wiki should remain.
  • Decide whether you need image handling.
  • Decide whether you need frontmatter and metadata conventions.
  • Decide whether you need search infrastructure.
  • Decide whether you want non-markdown outputs such as charts, canvases, or slide decks.
  • Decide which parts of this pattern are useful for your domain and ignore the rest.

Definition Of Done

  • You can add a source and the LLM updates the wiki without ad hoc instructions.
  • You can ask a question and get an answer grounded in the maintained wiki.
  • The wiki captures durable insights instead of losing them in chat.
  • index.md helps both you and the LLM navigate the knowledge base.
  • log.md gives a usable timeline of changes.
  • The schema is clear enough that future sessions can continue the workflow consistently.
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment