Karpathy Wiki Read

karpathy-wiki read

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."

The 6-step ladder

Every step is deterministic. There is no agent judgement at branch points; each branch is gated on a counted or boolean condition.

Step A.0: Unconfigured workspace

If no wiki root resolves because workspace routing is unconfigured, do not silently choose or initialize one. Skip the impossible schema/index reads and continue directly at Step F. After answering, follow the required capture flow. The capture skill is the sole owner of orphan preservation and the project, main, or both question. Do not add a second setup prompt here.

Malformed or incomplete configured routing is an error, not this branch.

Step A — Orient

Read these files in order:

1. <wiki>/schema.md — taxonomy, conventions, thresholds.
2. The relevant <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.

Step B — Count signal-matching candidates

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). For each term, count how many entries it substring-matches (case-insensitive) in the title, the one-liner, or the tag list on that line. A term that hits 6 or more entries is a common term (the Explore band). Other hitting terms are rare terms. Do not keep a denylist of tag names; frequency on this index is the only test.

Build the candidate set as follows:

• If there is at least one rare term, start with pages that match at least one rare term. Otherwise start with pages that match at least one common term.
• Then, for each common term, drop candidates that do not match it, unless that would leave zero pages; in that case skip that term.

A page matches a term when that term substring-hits its title, one-liner, or tag list.

Count the candidates. Branch on count:

• 0 candidates → Step F (cold-result path)
• 1-5 candidates → Step C (inline read)
• 6+ candidates → Step E (Explore subagent)

The threshold is 5/6, not your judgement. Do not "feel" your way to a different branch.

Step C — Inline read

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:

• YES → cite + answer (see "Cite contract" below). Done.
• NO → Step D (gap-fill via web search).

Step D — Gap-fill via web search

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:

1. Web search for the specific claim (use WebFetch or WebSearch).
2. Cite the web source(s) for the gap-filled claim.

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.