Agent skill

Using Karpathy Wiki

by toolboxmd in toolboxmd/karpathy-wiki

Auto-loaded by the SessionStart hook. An agent skill from toolboxmd/karpathy-wiki.

MITAuto-check passedKnowledge Management

Install Using Karpathy Wiki

skills CLI
$ npx skills add toolboxmd/karpathy-wiki --skill using-karpathy-wiki -a claude-code

Project install by default; add -g for ~/.claude/skills/.

GitHub CLI
$ gh skill install toolboxmd/karpathy-wiki using-karpathy-wiki --agent claude-code

Project scope by default; add --scope user for a personal install. Needs GitHub CLI 2.90.0 or later (public preview).

Manual copy
$ git clone --depth 1 https://github.com/toolboxmd/karpathy-wiki.git skills-src && mkdir -p .claude/skills && cp -r skills-src/skills/using-karpathy-wiki .claude/skills/using-karpathy-wiki && rm -rf skills-src

Use ~/.claude/skills/ instead of .claude/skills for a personal install. The folder must contain SKILL.md.

Claude Code skills documentation · loads skills from .claude/skills/

Facts

Skill name
using-karpathy-wiki
GitHub stars
105
Token cost
~2.2k tokens
SKILL.md length
1,197 words
Files
1
Skills in repo
5
Repo updated
First seen
Licence
MIT

At a glance

Auto-loaded by the SessionStart hook. An agent skill from toolboxmd/karpathy-wiki.

  • Tasks that involve LLM wikis
  • SKILL.md covers Announce when you act, Instruction priority, Iron laws and Orient before answering — the…, plus 5 more sections
  • Instructions only: no scripts, shell commands, URLs or credentials in SKILL.md

What it does

Using Karpathy Wiki is an agent skill from toolboxmd/karpathy-wiki. Auto-loaded by the SessionStart hook. Defines when to capture and routes the main agent or detached ingester to the appropriate on-demand skill.

Its SKILL.md is about 2.2k tokens, which your agent loads only when the skill is triggered. It is a single SKILL.md file with no bundled scripts.

It sits in Knowledge Management, covering LLM wikis. The repository describes itself as: Karpathy Wiki - Claude Code skills for building persistent, compounding knowledge bases. Based on Andrej Karpathy's LLM Wiki pattern. The licence is MIT.

When your agent uses it

  • Tasks that involve LLM wikis

Example prompts

  • “/using-karpathy-wiki”

What it can do on your machine

Read from SKILL.md and the folder at commit 67967f4. It shows what the files ask for, not the result of running them.

  • Tool permissions

    Pre-approves nothing: there is no allowed-tools line, so your agent's usual permission prompts apply.

    From allowed-tools in the SKILL.md frontmatter.

  • Runs code

    No scripts in the folder and no shell commands in SKILL.md (its code samples are bash).

    From the folder's file list and the shell code blocks in SKILL.md.

  • Network

    No URLs in SKILL.md.

    From URLs in SKILL.md, links to its own repository left out.

  • Credentials

    Names no API keys, tokens, secrets or passwords.

    From names ending in _API_KEY, _TOKEN, _SECRET, _KEY or _PASSWORD in SKILL.md.

Context cost

Using Karpathy Wiki loads about 2.2k tokens when it runs. Until then it costs about 41 tokens; SKILL.md has 1,197 words of instructions outside code blocks.

Always · name and description, kept in context so the agent knows when to use it
~41
When it runs · the whole SKILL.md, loaded when a task matches
~2.2k

Estimates: characters ÷ 4, the usual rule of thumb; real counts depend on the model's tokenizer. Scripts and assets cost tokens only if the agent reads them.

Safety

Auto-check passed

The automated check found no risky patterns in SKILL.md.

Automated static check — not a guarantee. Review scripts before installing. It scans the text of SKILL.md for risky patterns (piping downloads into a shell, reading credential files, hidden Unicode, destructive commands); files beside SKILL.md are not scanned.

SKILL.md

The full file from toolboxmd/karpathy-wiki at commit 67967f4, republished under its MIT licence (© toolboxmd). 1,197 words, ~2,195 tokens.

Download SKILL.mdSave it as .claude/skills/using-karpathy-wiki/SKILL.md (or your agent's skills folder).
name
using-karpathy-wiki
description
Auto-loaded by the SessionStart hook. Defines when to capture and routes the main agent or detached ingester to the appropriate on-demand skill.
<SUBAGENT-STOP>
If you were dispatched as a subagent for a task unrelated to wiki capture or ingest, skip this skill. The wiki rules apply to the main agent in the conversation; subagents doing isolated work (Explore, code-reviewer, codex:rescue, etc.) do not need to capture.
</SUBAGENT-STOP>

karpathy-wiki — loader (read once per session, applies for the whole session)

You are operating in a project that uses the karpathy-wiki plugin. Durable knowledge surfaced in this conversation must be captured to a wiki so future sessions can find it.

Announce when you act

When you write a capture, run an ingest, or answer from the wiki, prefix your reply with one line:

Using the karpathy-wiki skill to [capture this / ingest pending captures / answer from wiki].

This is the ONLY wiki-mechanics text the user sees. Do not narrate orientation, capture authoring, dispatch mechanics, or state-machine progress. Do all the wiki work silently after the announce line.

Instruction priority

If the user's CLAUDE.md / AGENTS.md or an explicit user instruction conflicts with the rules below, follow the user. The wiki rules override the default system prompt where they conflict, but never override the user's direct instructions.

Iron laws

NO WIKI WRITE IN THE FOREGROUND
NO PAGE EDIT WITHOUT READING THE PAGE FIRST
NO SKIPPING A CAPTURE BECAUSE "IT DOESN'T LOOK WIKI-SHAPED"
NO ANSWERING ANY USER QUESTION WITHOUT ORIENTING FIRST

Captures go to <wiki>/.wiki-pending/<timestamp>-<slug>.md. The bounded dispatcher assigns each capture to the configured provider/model profile and ingests it in the background.

Orient before answering — the read protocol

For ANY user message that is asking a QUESTION (vs requesting a code edit, file operation, or pure command), you must orient on the wiki BEFORE drafting the answer. Pre-classifying a question as "general knowledge" or "not wiki-relevant" is the forbidden rationalization Iron Rule 4 was written to forbid.

Cost framing. Orientation is a once-per-session investment:

  • First wiki-eligible question of the session: read <wiki>/schema.md and the relevant <wiki>/<category>/_index.md (or root <wiki>/index.md for cross-category questions). Two file reads, ~10-30 KB total.
  • Subsequent questions in the same session: the schema and index are already in your working memory; the marginal check is a candidate-count scan against the (already-read) _index.md. Near-zero cost.

Load karpathy-wiki-read/SKILL.md for the deterministic 6-step ladder (orient → count candidates → inline-read OR subagent OR web-search → cite). Do NOT skip the load because "this question seems trivial" — the iron rule is unconditional.

Resist-table for the read protocol (capture-side resist-table is below, after TRIGGER):

RationalizationReality
"This is general knowledge / I know this from training"Pre-classifying questions as wiki-irrelevant is the forbidden rationalization. The wiki's scope is whatever has been captured. You don't know without checking. Orient.
"This question is trivial; reading two files is overkill"Orientation is once per session. After the first orient, the marginal cost on the next question is near-zero (a re-scan of the already-read _index.md). The "trivial question" carve-out is the same drift that dropped the read protocol in v2.4.
"Pure syntax question, the wiki won't cover this"You cannot know without checking. If _index.md has zero matches, the read skill's Step F handles it (web search + capture the gap). The orient itself is what proves there are no candidates.
"User asked me to fix a bug, not answer a question"Code-edit requests don't trigger the read protocol; question-shaped requests do. If unsure (e.g. "why isn't this working" — could be a question OR a debug request), orient.
"I already oriented earlier this session, I'll skip this time"Correct! Re-use the schema and index you already have. Run only the candidate-count check on the new question's terms. The protocol's deterministic part (Step B) is the only thing that re-runs per question.

TRIGGER — when to capture

Write a capture when ANY of these fire (immediate-capture triggers):

  • A research subagent or research agent completes or returns a file with durable findings.
  • New factual information is found (web search, doc fetch, external fact surfaced in conversation).
  • Session resolves a confusion — something neither of you knew before.
  • A gotcha, quirk, or non-obvious behavior in a tool is discovered.
  • A pattern is validated (approaches compared, one picked, with reasons).
  • An architectural decision is made with explicit rationale.
  • User pastes a URL or document for study.
  • <wiki>/inbox/ or <wiki>/raw/ has unprocessed files.
  • User says "add to wiki" / "remember this" / "wiki it" / "save this".
  • Two claims contradict each other.

Also TRIGGER (orientation + citation): the user asks "what do we know about X" / "how do we handle Y" / "what did we decide about Z" / "have we seen this before" — or any question the wiki might cover.

Show full SKILL.md (480 more words)Show less

SKIP — when not to capture

  • Routine file edits, syntax lookups, one-off debugging with trivial root causes.
  • Time-sensitive data that must be fetched fresh.
  • Questions clearly outside any wiki's scope.

Do NOT skip based on tone or shape. "This looks like casual chat", "there's no code here", "this isn't a wiki context" are forbidden rationalizations. If new factual information appeared, capture. Tone is not the trigger.

Resist-table for capture triggers. When you're about to skip a capture, check these red flags. Skipped captures are invisible — the user cannot observe them, so the discipline is yours alone.

RationalizationReality
"The user will remember this"The user will not remember. That's the whole point.
"It's too trivial for the wiki"If a TRIGGER fires, capture. Lint and the ingester filter noise later.
"I'll capture it later"Later means never. Capture now — invocation is milliseconds.
"I'm in the middle of another task"Capture is non-blocking; bin/wiki capture writes a file and returns. The ingester runs detached.
"The user didn't ask me to save this"Triggers fire automatically. No explicit user request is required — that's the loader's contract.
"I don't have a memory tool available"This skill IS the memory tool. The loader's presence is the trigger.
"The file is already in a good place"Filing ≠ capturing. Location is not organization. The capture extracts the durable concept from the file's context; the file alone doesn't.

What to do when a trigger fires

For chat-driven captures (the conversation IS the source, or you have a file alongside the conversation): load karpathy-wiki-capture/SKILL.md and follow it. The capture is written via bin/wiki capture (a subcommand of the existing bin/wiki CLI).

For research subagent reports (a file is the source): the agent's job is one command, NOT a hand-written capture body:

bash
mv <subagent-report-path> <wiki>/inbox/<basename>
wiki ingest-now <wiki>          # or wait for next SessionStart

For ANY user question (including "what does the wiki know about X" questions, but also any other question per Iron Rule 4): load karpathy-wiki-read/SKILL.md and run the deterministic 6-step ladder. Do NOT load karpathy-wiki-ingest/SKILL.md — that's for the detached runtime ingester only and contains write-side machinery the main agent never needs.

Mode change

If the user asks to change the wiki mode for this directory (phrases like "use project wiki here," "main only," "switch to both," "fork to main"), run:

bash
wiki use project|main|both

Confirm the change in one line; do not over-narrate.

What's NOT in this loader

The full operational details — capture format, body-size floors, dispatch mechanics, ingester orientation, page format, manifest protocol, commit conventions — live in the on-demand skills:

  • skills/karpathy-wiki-capture/SKILL.md — load when you are about to write a capture.
  • skills/karpathy-wiki-read/SKILL.md — load when you are about to answer a user question (Iron Rule 4).
  • skills/karpathy-wiki-ingest/SKILL.md — loaded by the detached ingester via its provider prompt; the main agent never reads this.
  • skills/karpathy-wiki-doctor/SKILL.md — detached census; wiki doctor launches it. The main agent never reads this.

If you do not know what to do at a particular step, load the on-demand skill — do not invent.

© toolboxmd, MIT. Rendered from Markdown: HTML in the file is shown as text, images as links, and headings moved down two levels. Raw file

Files

Just SKILL.md in skills/using-karpathy-wiki of toolboxmd/karpathy-wiki.

Open the folder on GitHubat commit 67967f4

Compare with similar skills

Using Karpathy Wiki next to the 5 skills that share the most tags, products or categories with it. Stars are the repository's; “used in” counts other GitHub owners with a copy.

Using Karpathy Wiki compared with similar skills
SkillStarsUsed inTokensAuto-checkLicenceRepo updated
Using Karpathy Wiki this skilltoolboxmd/karpathy-wiki105—~2.2kAutomated safety check: PassMIT
Karpathy LLM WikiAstro-Han/karpathy-llm-wiki2.4k—~3.6kAutomated safety check: PassMIT
LLM Wiki Knowledge GraphEgonex-AI/Understand-Anything86k1 repos~1.5kAutomated safety check: PassMIT
Hermes History IngestAr9av/obsidian-wiki3.5k1 repos~2.2kAutomated safety check: NotesMIT
LLM Wikilewislulu/llm-wiki-skill655—~3.7kAutomated safety check: PassNone
Wiki Builderrohitg00/pro-workflow2.9k—~1kAutomated safety check: PassNone

Similar skills

  • Karpathy LLM Wiki

    Astro-Han/karpathy-llm-wiki

    A skill your agent uses when building or maintaining a personal LLM-powered knowledge base.

    2.4k GitHub stars~3.6k tokensUpdated 2 mo ago
    Knowledge ManagementAuto-check passed
  • LLM Wiki Knowledge Graph

    Egonex-AI/Understand-Anything

    Detects a Karpathy-pattern LLM wiki and builds an interactive knowledge graph with entities, implicit relationships and topic clusters.

    86k GitHub starsUsed in 1 repo~1.5k tokens
    Knowledge ManagementAuto-check passed
  • Hermes History Ingest

    Ar9av/obsidian-wiki

    Ingest Hermes agent history into Obsidian as distilled knowledge.

    3.5k GitHub starsUsed in 1 repo~2.2k tokens
    Knowledge ManagementAuto-check: notes
  • LLM Wiki

    lewislulu/llm-wiki-skill

    Build and maintain a Karpathy-style LLM knowledge base — a self-compiling Obsidian markdown wiki where an Agent ingests raw sources, compiles cross-linked concept/entity/summary pages, answers…

    655 GitHub stars~3.7k tokensUpdated 5 mo ago
    Knowledge ManagementAuto-check passed
  • Wiki Builder

    rohitg00/pro-workflow

    Start, structure, and grow a persistent research wiki indexed in pro-workflow's SQLite knowledge base.

    2.9k GitHub stars~1k tokensUpdated 9 days ago
    Knowledge ManagementAuto-check passed
  • Wiki Query

    rohitg00/pro-workflow

    Query pro-workflow wikis via SQLite FTS5 BM25 retrieval. An agent skill from rohitg00/pro-workflow.

    2.9k GitHub starsUsed in 1 repo~554 tokens
    Knowledge ManagementAuto-check passed

More from toolboxmd/karpathy-wiki

  • Karpathy Wiki Capture

    toolboxmd/karpathy-wiki

    Capture-authoring protocol for the main agent. An agent skill from toolboxmd/karpathy-wiki.

    105 GitHub stars~2k tokensUpdated yesterday
    Auto-check passed
  • Karpathy Wiki Doctor

    toolboxmd/karpathy-wiki

    Detached doctor only. An agent skill from toolboxmd/karpathy-wiki.

    105 GitHub stars~489 tokensUpdated yesterday
    Auto-check passed
  • Karpathy Wiki Ingest

    toolboxmd/karpathy-wiki

    Detached ingester only. An agent skill from toolboxmd/karpathy-wiki.

    105 GitHub stars~6.8k tokensUpdated yesterday
    Auto-check passed
  • Karpathy Wiki Read

    toolboxmd/karpathy-wiki

    Read protocol for the main agent. An agent skill from toolboxmd/karpathy-wiki.

    105 GitHub stars~1.9k tokensUpdated yesterday
    Auto-check passed

Questions about Using Karpathy Wiki

What does Using Karpathy Wiki do?

Auto-loaded by the SessionStart hook. An agent skill from toolboxmd/karpathy-wiki. Using Karpathy Wiki is an agent skill from toolboxmd/karpathy-wiki. Auto-loaded by the SessionStart hook.

When should I use Using Karpathy Wiki?

Using Karpathy Wiki fits situations like: tasks that involve LLM wikis.

How do I install Using Karpathy Wiki in Claude Code?

Run `npx skills add toolboxmd/karpathy-wiki --skill using-karpathy-wiki -a claude-code`. Or copy the skill folder (skills/using-karpathy-wiki in toolboxmd/karpathy-wiki) into .claude/skills/using-karpathy-wiki in your project. Claude Code loads it when a task matches its description.

How do I install Using Karpathy Wiki in Codex?

Run `npx skills add toolboxmd/karpathy-wiki --skill using-karpathy-wiki -a codex`. Or copy the skill folder (skills/using-karpathy-wiki in toolboxmd/karpathy-wiki) into .agents/skills/using-karpathy-wiki in your project. Codex loads it when a task matches its description.

Can I use Using Karpathy Wiki in Cursor, Gemini CLI or GitHub Copilot?

Cursor, Gemini CLI, GitHub Copilot and OpenCode also load SKILL.md folders. With the skills CLI, run `npx skills add toolboxmd/karpathy-wiki --skill using-karpathy-wiki -a cursor` (or -a gemini-cli, github-copilot or opencode for the others). To copy it by hand, put the folder in .cursor/skills/using-karpathy-wiki, .gemini/skills/using-karpathy-wiki, .github/skills/using-karpathy-wiki and .opencode/skills/using-karpathy-wiki in your project.

What does Using Karpathy Wiki need to run?

SKILL.md names no scripts, command-line tools or credentials: Using Karpathy Wiki is instructions for the agent only.

Does Using Karpathy Wiki access the network?

SKILL.md contains no URLs. Any network use would come from the scripts or tools the agent runs. This is read from the text; nothing was executed.

Is Using Karpathy Wiki safe to install?

Our automated static check of SKILL.md found no risky patterns, such as piping downloads into a shell, reading credential files or hidden Unicode. It is not a guarantee. Review the folder before installing.

What licence does Using Karpathy Wiki use?

Using Karpathy Wiki is published under the MIT licence (the repository's licence). It allows redistribution, so the full SKILL.md is shown on this page.

How many tokens does Using Karpathy Wiki use?

About 2.2k tokens (SKILL.md is roughly 8.8k characters). Agents keep only the skill's name and description in context until a task matches; then they load SKILL.md in full.

What are the alternatives to Using Karpathy Wiki?

Skills that share tags, products or a category with Using Karpathy Wiki: Karpathy LLM Wiki (Astro-Han/karpathy-llm-wiki, 2.4k stars), LLM Wiki Knowledge Graph (Egonex-AI/Understand-Anything, 86k stars), Hermes History Ingest (Ar9av/obsidian-wiki, 3.5k stars) and LLM Wiki (lewislulu/llm-wiki-skill, 655 stars). The comparison table on this page puts their stars, adoption, token cost, safety result and licence side by side.

Who maintains Using Karpathy Wiki?

toolboxmd (a GitHub organization) maintains it in toolboxmd/karpathy-wiki, which has 105 GitHub stars. The repository holds 5 skills in this directory. The repository was last updated on October 7, 2026.

Source: toolboxmd/karpathy-wiki on GitHub. Facts on this page come from the repository at the commit we read; the author's words are quoted as theirs.