Records hypotheses, evidence and decisions when proposing, running or evaluating forecast experiments, including backtests and feature or model comparisons. Excludes routine tests and unchanged forecast operation.
한국어 요약은 준비 중입니다. 원문 설명을 표시하고 있습니다.
Build and maintain a persistent, LLM-maintained markdown wiki following Andrej Karpathy's LLM Wiki pattern — the agent incrementally compiles raw sources into an interlinked knowledge base instead of re-deriving answers from scratch. Use this whenever the user wants to ingest a source (an article, paper, transcript, chapter, meeting notes), file knowledge or conversation findings into the wiki, query the accumulated knowledge base, lint or health-check the wiki, or mentions "the wiki", "knowledge base", "ingest this", "file this", "what do we know about X", or accumulating notes on a topic over time — even if they don't say "LLM Wiki" explicitly.
한국어 요약은 준비 중입니다. 원문 설명을 표시하고 있습니다.
설치 방법 보기설치하기 전에 에이전트에게 전달되는 지시 내용을 확인할 수 있습니다.
This skill instantiates Andrej Karpathy's LLM Wiki pattern. The verbatim source of the idea is bundled at references/karpathy-llm-wiki.md — read it in full the first time you use this skill in a session, and whenever you need to make a design decision the summary below doesn't cover. It is the authority; this file is only the operational digest.
Instead of retrieving from raw documents on every question (RAG), the agent incrementally builds and maintains a persistent wiki: a structured, interlinked collection of markdown files sitting between the user and the raw sources. Knowledge is compiled once and kept current, not re-derived per query. The wiki is a compounding artifact: cross-references already exist, contradictions are already flagged, the synthesis already reflects everything read so far.
The human curates sources, directs the analysis, and asks the questions. The agent does all the bookkeeping: summarizing, cross-referencing, filing, updating, flagging contradictions. The human never (or rarely) writes wiki pages by hand.
Ingest — a new source lands in the raw collection: read it, discuss key takeaways with the user, write a summary page, update the index, update every entity/concept page it touches (a single source may touch 10–15 pages), and append a log entry. Prefer one source at a time with the user involved, unless they ask for a batch.
Query — answer questions from the wiki: read index.md first to find
relevant pages, drill in, synthesize with citations to wiki pages. When an
answer is valuable — a comparison, an analysis, a discovered connection — file
it back into the wiki as a new page so explorations compound instead of
vanishing into chat history.
Lint — periodically health-check: contradictions between pages, claims superseded by newer sources, orphan pages, concepts mentioned but lacking a page, missing cross-references, data gaps worth a web search. Suggest new questions and sources.
index.md — content catalog: every page with a link and one-line
summary, organized by category. Update on every ingest; read first on every
query. Works to ~hundreds of pages without any search infrastructure.log.md — append-only chronology of ingests/queries/lints. Start each
entry with a consistent, grep-able prefix:
## [2026-04-02] ingest | Article TitleConciseness is of utmost importance. The human's attention is the bottleneck: wiki pages exist to be read fast, so write only what carries knowledge — no filler, no restating what a linked page or the README already says.
The wiki lives in wiki/ at the repo top level and is committed to git.
Before writing anything into it, read wiki/AGENTS.md and honor its rules —
most importantly its content policy: the wiki holds knowledge about this
codebase and project only, never personal information about the user. Raw
sources, when the wiki needs them, go in wiki/raw/ (immutable — read, never
modify). The special files wiki/index.md and wiki/log.md exist and must be
kept current as described above.
The pattern is intentionally abstract — directory layout, page formats, and
tooling depend on the domain. When setting up a new wiki with the user, decide
together: where raw sources and wiki pages live, what page categories exist,
what frontmatter/tags to use, and which optional pieces (search tooling like
qmd, Obsidian integration, image handling, output formats like Marp decks)
are actually needed. Everything optional stays out until the need is real.
Record the decisions in the wiki's own schema/conventions file so future
sessions inherit them.
아직 리뷰가 없습니다. 사용해 본 소감을 남겨 주세요.
로그인 없이 쓸 수 있습니다. 사용 후기, 질문, 작동하지 않은 환경 등을 편하게 남겨 주세요.
아직 댓글이 없습니다. 첫 댓글을 남겨 보세요.
Records hypotheses, evidence and decisions when proposing, running or evaluating forecast experiments, including backtests and feature or model comparisons. Excludes routine tests and unchanged forecast operation.
한국어 요약은 준비 중입니다. 원문 설명을 표시하고 있습니다.
Diagnose GitHub CLI hiccups caused by Codex sandbox constraints; not a general GitHub workflow.
한국어 요약은 준비 중입니다. 원문 설명을 표시하고 있습니다.
Review a design artifact (spec, plan, PR packet) with fresh, memoryless subagents. Invoke when the user asks — "goldfish this", "fresh eyes", "tear this apart" — or after suggesting it for a spec that has grown load-bearing and the user agrees.
한국어 요약은 준비 중입니다. 원문 설명을 표시하고 있습니다.
Interrogate the user before implementing a non-trivial or ambiguous ask, converging on shared intent. Use when the user says "grill me", when an ask admits multiple interpretations, or before any sizable piece of work whose intent isn't fully clear.
한국어 요약은 준비 중입니다. 원문 설명을 표시하고 있습니다.
Launches and supervises long-running commands — solves, workflows, downloads, builds — detached with a log, PID and exit sentinel, then tells stuck from slow and kills verifiably. Use before running anything that may exceed a minute, for nohup or background launches, and when a run hangs, shows no output, or timed out.
한국어 요약은 준비 중입니다. 원문 설명을 표시하고 있습니다.
Guides writing and editing Mermaid diagrams in GitHub-rendered Markdown so they remain readable across themes and preserve placeholder labels.
한국어 요약은 준비 중입니다. 원문 설명을 표시하고 있습니다.