跳至正文
OpenAgentSkill
浏览已开放 · 目录记录日期 2026-11-21 · 自动更新和账户服务暂时关闭。
搜索与研究

karpathy-wiki-read

Read protocol for the main agent. Load on demand when answering ANY user question (per Iron Rule 4 in `using-karpathy-wiki/SKILL.md`). Defines the deterministic 6-step orientation ladder for finding wiki coverage of a question, when to inline-read vs spawn an Explore subagent vs fall through to web search, and the cite contract every wiki-grounded answer must satisfy.

toolboxmd · 已有 AI 审核记录

安装与使用

运行命令前先阅读来源与依赖说明。复制命令不会自动安装或执行。

npx skills add toolboxmd/karpathy-wiki --skill karpathy-wiki-read

技能说明


name: karpathy-wiki-read description: | Read protocol for the main agent. Load on demand when answering ANY user question (per Iron Rule 4 in using-karpathy-wiki/SKILL.md). Defines the deterministic 6-step orientation ladder for finding wiki coverage of a question, when to inline-read vs spawn an Explore subagent vs fall through to web search, and the cite contract every wiki-grounded answer must satisfy.

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 — 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). A page is a candidate if ANY signal term:

  • Substring-matches its title (case-insensitive), OR
  • Matches a tag (exact, case-insensitive), OR
  • Appears in its one-line summary.

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.

Step E — Explore subagent

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:

  • If it returned a synthesis with cited pages → use the synthesis as the answer's basis. Add conversational framing as needed. Cite the pages the subagent cited.
  • If it reported "wiki does not cover" → Step D (web search + capture the gap).

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.

Step F — Cold result (no candidates)

Zero signal matches in _index.md. The wiki has no coverage for this question.

  1. Web search for the answer (WebFetch or WebSearch).
  2. Cite the web source(s) in the answer.
  3. Always write a capture so the next session is not cold for this topic. The capture's title is the question; the body documents what was found and where (the URLs you searched). Use 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.

Cite contract

Every wiki-grounded answer must include citations the user can verify. Format:

  • For each wiki page you drew from: cite as <wiki>/<category>/<page>.md followed by a brief relevance note OR a short quote (≤2 lines).
  • For each web source you drew from: cite as [<title>](<url>).
  • Citations appear inline near the claim they support, not in a footer block, so the user can check each claim against its source.

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.

What goes in this skill — and what doesn't

This skill covers the read-from-wiki protocol only. It does NOT cover:

  • Capture authoring (when a question reveals a wiki gap, the capture flow lives in `karpat
阅读完整来源 ↗