Help us improve
Share bugs, ideas, or general feedback.
From karpathy-wiki
Implements 6-step deterministic ladder to check Karpathy wiki coverage for any user question: orient via schema/index, count signal matches, inline-read candidates or spawn Explore subagent or web search.
npx claudepluginhub toolboxmd/karpathy-wiki --plugin karpathy-wikiHow this skill is triggered — by the user, by Claude, or both
Slash command
/karpathy-wiki:karpathy-wiki-readThe summary Claude sees in its skill listing — used to decide when to auto-load this skill
You loaded this skill because the user asked a question (per Iron Rule 4 in the loader: NO ANSWERING ANY USER QUESTION WITHOUT ORIENTING FIRST). Your job is to determine whether the wiki covers the answer, retrieve the coverage, and either cite-and-answer from the wiki or fall through to web search — without ever pre-classifying the question as "wiki-irrelevant."
Answers domain questions by reading and synthesizing project wiki pages, citing sources with wikilinks. Flags gaps and contradictions.
Search and query the wiki — get cited answers, with automatic research when knowledge is missing. Three depth modes. Use when: 'check wiki', 'what do we know about', 'look up', 'find in wiki', 'search wiki', 'search for', 'find information about', 'do we have notes on', 'wiki context', 'existing knowledge', 'research', 'look into', 'find out about', 'investigate', 'what is', 'how does X work'.
Queries personal wiki built with wiki-init/wiki-ingest: reads index/pages, synthesizes cited answers for facts/comparisons/how-tos, flags gaps, offers to save as new pages.
Share bugs, ideas, or general feedback.
You loaded this skill because the user asked a question (per Iron Rule 4 in the loader: NO ANSWERING ANY USER QUESTION WITHOUT ORIENTING FIRST). Your job is to determine whether the wiki covers the answer, retrieve the coverage, and either cite-and-answer from the wiki or fall through to web search — without ever pre-classifying the question as "wiki-irrelevant."
Every step is deterministic. There is no agent judgement at branch points; each branch is gated on a counted or boolean condition.
Read these files in order:
<wiki>/schema.md — taxonomy, conventions, thresholds.<wiki>/<category>/_index.md for the question's apparent topic. If the question crosses categories or you cannot tell which category applies, read <wiki>/index.md (the root MOC).If you have already oriented earlier in this session, skip to Step B — the schema and index content are already in your working memory. Orientation is once per session.
Extract the question's signal terms: meaningful nouns, proper-noun phrases, technical terms, version numbers, tool names. Skip stopwords ("the", "what", "how", "do", "is").
Walk the relevant _index.md (already in memory from Step A). A page is a candidate if ANY signal term:
Count the candidates. Branch on count:
The threshold is 5/6, not your judgement. Do not "feel" your way to a different branch.
Read all candidate pages in full. After reading, ask exactly:
Does the union of these pages contain every claim my answer would make?
Branch on the answer:
The wiki covered part of the answer; the rest needs fresh information. For each claim your answer would make that the wiki did NOT cover:
WebFetch or WebSearch).Compose the answer using BOTH the wiki citations (for what it covered) and the web citations (for what it didn't).
After answering, always write a capture noting the gap. The wiki should grow toward questions it failed to answer fully. The capture's title is the question; the body documents what the wiki had and what was missing. Use karpathy-wiki-capture/SKILL.md for the capture authoring procedure.
Spawn an Explore subagent with this prompt shape:
Question:
<the user's question, verbatim>Wiki path:
<wiki absolute path>Run the orient procedure (read schema.md and the relevant _index.md), identify candidate pages by signal-term match, read all candidates in full (no cap on page count — your context is isolated), and return a synthesis.
The synthesis must:
- Cite every page it draws from (path + one-line relevance note per page).
- Cover the question completely (no truncation for brevity if the answer is genuinely long).
- Be as terse as possible while preserving every wiki-specific claim, decision, and contradiction. Terseness is a property of the writing, not a target word count.
- If the wiki does not cover the question, say so explicitly and return an empty page-list. Do NOT attempt to fall through to web search yourself; the main agent handles that.
When the subagent returns:
The subagent dispatch maps to the patterns in your ~/.claude/CLAUDE.md Task Delegation rule: this is a context-isolation case where the main agent needs the synthesis, not the page contents.
Zero signal matches in _index.md. The wiki has no coverage for this question.
WebFetch or WebSearch).karpathy-wiki-capture/SKILL.md.A cold result is not a failure — it is the wiki telling you it has a gap, and you closing the gap.
Every wiki-grounded answer must include citations the user can verify. Format:
<wiki>/<category>/<page>.md followed by a brief relevance note OR a short quote (≤2 lines).[<title>](<url>).If your answer makes a claim with no citation, that claim came from training data — flag it explicitly: "(from training data; not in wiki)." This is the only case where uncited claims are allowed, and they must be flagged.
This skill covers the read-from-wiki protocol only. It does NOT cover:
karpathy-wiki-capture/SKILL.md).karpathy-wiki-ingest/SKILL.md and apply only to the spawned ingester.using-karpathy-wiki/SKILL.md).If you find yourself needing capture or ingest mechanics while running this protocol, load the appropriate sibling skill — do not improvise.