Build a Knowledge Base Claude Can Query

The Problem

Every new conversation starts from zero. You re-explain your business, your processes, your decisions — over and over.

The Method

A structured wiki of summaries, entity pages, and concept pages that Claude navigates like an index — reading only what's relevant.

The Result

80–85% fewer tokens per query. Persistent knowledge that builds over time. Answers with citations instead of hallucinated guesses.

Covered in Episode 34 — Token Limits, Wikis, and How to Query Your Knowledge Base · Karpathy's original gist

Tell us about your setup

Answer three quick questions and we'll show you the right prompts for your situation.

How do you access Claude?

Where do you keep your notes and files?

What do you mainly want to capture?

YOUR SETUP

Full four-skill suite

Claude Code + Obsidian gives you the complete workflow — automated file creation, index management, and cross-linking. Copy each prompt and use it in a Claude Code conversation with your vault open.

wiki-bootstrap — seed your wiki from existing content

You are helping me build a personal knowledge wiki.

My notes are stored at: [YOUR FOLDER PATH OR VAULT PATH]

Please:
1. Scan the folder for existing markdown files — procedures, notes, resources, episode notes, anything substantial
2. Create a Wiki/ subfolder with: Wiki/pages/sources/, Wiki/pages/entities/, Wiki/pages/concepts/
3. For each significant file, create a source summary page in Wiki/pages/sources/ with: title, type, key claims (3–5 bullets), entities mentioned, and concepts covered
4. Create stub pages for frequently mentioned entities (people, tools, companies) and concepts
5. Build Wiki/index.md as a searchable index with links to all pages
6. Report back with what you built and what's in the index

Use [[wikilinks]] for all internal references. Use kebab-case filenames. Every page needs a type field (source, entity, or concept) and a status field (complete or stub).

wiki-ingest — add a new source to your wiki

You are helping me add a new source to my personal knowledge wiki.

My wiki is at: [YOUR FOLDER PATH]/Wiki/

Source to ingest: [PASTE A URL, FILE PATH, OR THE CONTENT DIRECTLY]

Please:
1. Read or fetch the source content
2. Create a source summary page in Wiki/pages/sources/ with: title, author, date, key claims (3–5 bullets), entities mentioned, concepts covered, and one notable quote
3. Check Wiki/index.md and create or update entity pages for any new people, tools, or companies mentioned
4. Create or update concept pages for any significant ideas introduced
5. Add an entry to Wiki/index.md
6. Report what was created and flag any stubs that need filling out

wiki-query — ask a question across your wiki

You are helping me query my personal knowledge wiki.

My wiki is at: [YOUR FOLDER PATH]/Wiki/

My question: [YOUR QUESTION]

Please:
1. Read Wiki/index.md to identify relevant pages
2. Read the relevant source summaries and entity or concept pages
3. Synthesise an answer using only what's in the wiki, citing which pages you used
4. Flag anything you couldn't answer from the wiki alone
5. Suggest related pages I might want to read next

wiki-lint — health check your wiki

You are running a health check on my personal knowledge wiki.

My wiki is at: [YOUR FOLDER PATH]/Wiki/

Please:
1. Check Wiki/index.md for any listed pages that don't exist (dead links)
2. Find pages not listed in the index (unlisted pages)
3. Find orphan pages with no inbound links from other pages
4. List stub pages sorted by how many other pages reference them (highest priority first)
5. Flag any pages missing required frontmatter fields (type, status)
6. Suggest 3 quick wins — the highest-value missing pages to create next

Recommended order

Run wiki-bootstrap once to seed from existing content. Use wiki-ingest whenever you want to add something new. Run wiki-lint every few weeks to keep the connections clean.

What each skill does

wiki-bootstrap

A one-time operation that scans your existing content — procedures, notes, resources — and builds the initial wiki structure. Run it once to seed everything you already have.

wiki-ingest

Adds a new source to your wiki. Give it a URL, a file path, or raw content. It creates a summary page, updates entity pages, and links everything together.

wiki-query

Searches your wiki and synthesises an answer from multiple sources — with citations. Reads the index first, then drills into only what's relevant. Token-efficient by design.

wiki-lint

Runs a health check. Finds orphan pages, broken links, stubs with high demand, and missing frontmatter. Suggests the highest-value gaps to fill next.

What wiki-ingest produces

Raw input — a YouTube transcript excerpt

...what Karpathy basically said was you create stub pages for everything and then the AI navigates the index first before it ever hits the full file. It's basically a funnel. You query a little bit, get some direction, go deeper, get more information...

Structured output — saved to your wiki

## Karpathy Wiki Method

**Source:** YouTube · @karpathy

**Type:** source | **Status:** complete

Key Takeaways

- Index-first retrieval cuts token use 80–85%

- Stubs act as pointers, not full pages

- Graph view reveals hidden connections

Entities

[[andrej-karpathy]] · [[obsidian]] · [[claude]]

RELATED EPISODE

EP034 — Token Limits, Wikis, and How to Query Your Knowledge Base

Where the wiki methodology came from, why token costs matter for solo operators, and how it connects to building target operating models and agent teams.

Listen now

Want the full system?

The wiki skills are part of a broader setup for solo operators who want Claude running as a persistent, context-aware partner. Join Unleash AI to see how it all fits together.

Join Unleash AI All Resources

Build a Knowledge Base Claude Can Query

Tell us about your setup

Full four-skill suite

Recommended order

Two-prompt workflow

Want the full automated setup?

Start here

Recommended next steps

What each skill does

What wiki-ingest produces

EP034 — Token Limits, Wikis, and How to Query Your Knowledge Base

Want the full system?