菜单
Answer a question against the Loop Library wiki — read the index first, drill into relevant pages, synthesize with citations, and file genuinely new findings back as wiki pages. Use when the user asks a question about loops, patterns, or concepts that the knowledge base should answer.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
Keep this project's graph lean and trustworthy — merge duplicates, retire stale notes (respecting task hierarchy), resolve contradictions, fix links, resolve entities, flag type drift, re-derive the per-entity timelines, regenerate the index-views, and surface stale tasks. Suggest-only; asks before deleting. Use periodically, before an important piece of work, or when the graph feels noisy. (brainkit's name for Karpathy's "lint".)
Capture new material into this project's graph as a typed entity — resolve what it mentions to canonical notes (with you, at capture), expand its shorthand and link it, preserve the raw, append to any touched entity's timeline, and extract the reusable knowledge and tasks. Use whenever the user shares or refers to something the brain should remember — a call, a meeting note, a doc, a decision, a person, a task, a link.
Put this project's graph to work toward the deliverable — answer a grounded question, advance a work/ deliverable with citations, run gap analysis ("what's missing for this?"), or read the task and reading-queue views — and file genuinely new conclusions back. Use when the user asks something the brain should know, asks to move a deliverable forward, or asks what's open or unprocessed.
Curate upstream improvements into this kit — a research task, not a script. Read canonical brainkit upstream, reason about what genuinely improved, and pull in only the changes you and the user find compelling, preserving local tweaks. Suggest-only; never bulk-overwrites. Use when the user wants to refresh the kit's machinery from origin.
Curate upstream improvements into this kit — a research task, not a script. Read canonical karpathy-core upstream, reason about what genuinely improved, and pull in only the changes you and the user find compelling, preserving local tweaks. Suggest-only; never bulk-overwrites. Use when the user wants to refresh the kit's machinery from origin.
Curate upstream improvements into this kit — a research task, not a script. Read canonical loopkit upstream, reason about what genuinely improved, and pull in only the changes you and the user find compelling, preserving local tweaks. Suggest-only; never bulk-overwrites. Use when the user wants to refresh the kit's machinery from origin.
| name | query |
| description | Answer a question against the Loop Library wiki — read the index first, drill into relevant pages, synthesize with citations, and file genuinely new findings back as wiki pages. Use when the user asks a question about loops, patterns, or concepts that the knowledge base should answer. |
Karpathy's second operation: answer from the compiled wiki, then file good answers back so explorations compound. Read CLAUDE.md §3 first.
wiki/index.md to find candidate pages. Never blind-scan directories.related: and in-body /links one level at a time to traverse the graph./sources/ concepts you used. Distinguish what the wiki knows (extracted) from what you're inferring — don't present synthesis as established fact.wiki/ page (type: Comparison / Concept / Synthesis) with full frontmatter + # Citations, update the relevant index.md, and append a query entry to wiki/log.md: - query | <question> — filed as /path.md.scripts/conformance.sh.If the wiki can't answer, say so plainly and propose the source(s) to ingest next — don't fabricate. A missing page is "not-yet-written knowledge," not an error.
When a follow-up needs the user to choose — which output form, which of several readings to pursue, whether to file a new page — ask with the structured AskUserQuestion prompt (concrete options), not a free-text question.