「reference」の検索結果

Launch Marcus multi-agent experiments with independent CLI agents in tmux panes. Use when the user says "build X with Marcus", "run a Marcus experiment", "spawn N agents with Marcus", "/marcus", or references the Marcus multi-agent framework. This skill handles experiment setup, config generation, and spawning independent agent CLI processes that coordinate via the Marcus MCP server. Each agent runs in its own tmux pane with full autonomy. NOT for ClawTeam — this is specifically for the Marcus MCP-based multi-agent system. NOTE: "Marcus" is also the name of the MCP server that agents use for coordination. The /marcus skill launches experiments that USE the Marcus MCP server — they are complementary, not competing. The MCP server must be running before invoking this skill. The Marcus MCP server is agent-agnostic; this skill supports both Anthropic's claude CLI and OpenAI's codex CLI as agent harnesses (selected via --harness).

日本語の概要は準備中です。原文の説明を表示しています。

Launch Marcus multi-agent experiments with independent Claude agents in tmux panes. Use when the user says "build X with Marcus", "run a Marcus experiment", "spawn N agents with Marcus", "/marcus", or references the Marcus multi-agent framework. This skill handles experiment setup, config generation, and spawning independent Claude CLI agents that coordinate via the Marcus MCP server. Each agent runs in its own tmux pane with full autonomy. NOT for ClawTeam — this is specifically for the Marcus MCP-based multi-agent system. NOTE: "Marcus" is also the name of the MCP server that agents use for coordination. The /marcus skill launches experiments that USE the Marcus MCP server — they are complementary, not competing. The MCP server must be running before invoking this skill.

日本語の概要は準備中です。原文の説明を表示しています。

Launch Marcus multi-agent experiments with independent CLI agents in tmux panes. Use when the user says "build X with Marcus", "run a Marcus experiment", "spawn N agents with Marcus", "/marcus", or references the Marcus multi-agent framework. This skill handles experiment setup, config generation, and spawning independent agent CLI processes that coordinate via the Marcus MCP server. Each agent runs in its own tmux pane with full autonomy. NOT for ClawTeam — this is specifically for the Marcus MCP-based multi-agent system. NOTE: "Marcus" is also the name of the MCP server that agents use for coordination. The /marcus skill launches experiments that USE the Marcus MCP server — they are complementary, not competing. The MCP server must be running before invoking this skill. The Marcus MCP server is agent-agnostic; this skill supports Anthropic's claude CLI, OpenAI's codex CLI, and Google's gemini CLI as agent harnesses (selected via --harness).

日本語の概要は準備中です。原文の説明を表示しています。

Create or review dygo Fixtures and explicit Patches for reference data, setup Records, backfills, renames, and unsafe lifecycle transitions. Use when metadata sync alone cannot express the required data change.

日本語の概要は準備中です。原文の説明を表示しています。

Write or update dygo instructions, reference documentation, task descriptions, CLI help, metadata examples, or architecture notes. Use when documentation quality and contract accuracy are central to the task.

日本語の概要は準備中です。原文の説明を表示しています。

Capture and organize knowledge in Notion—research findings, best practices, reference materials. Build searchable knowledge base for your team and future self.

日本語の概要は準備中です。原文の説明を表示しています。

Create distinctive, production-ready frontend interfaces with high visual quality. Use this skill whenever the user asks to build, design, or improve any web UI — including landing pages, dashboards, components, app screens, forms, or interactive widgets. Also trigger when the user says "make this look better", "redesign this", "add polish", "build a UI", "create a page", "style this", or references HTML/CSS/JS, React, Vue, Svelte, or any frontend framework. If the request involves any visible output rendered in a browser, use this skill — even if the user hasn't explicitly said "design." This skill produces intentional, memorable interfaces that avoid generic AI patterns.

日本語の概要は準備中です。原文の説明を表示しています。

Use when the user asks to run Codex CLI (codex exec, codex resume) or references OpenAI Codex for code analysis, refactoring, or automated editing. Uses GPT-5.2 by default for state-of-the-art software engineering.

日本語の概要は準備中です。原文の説明を表示しています。

Reference for MCP Jupyter tools (kernel management, cell execution, Papermill). Use when executing notebooks, managing kernels, or running code interactively via MCP.

日本語の概要は準備中です。原文の説明を表示しています。

Reference for notebook manipulation scripts (notebook_helpers.py, notebook_tools.py). Use when working with Jupyter notebooks, analyzing structure, executing cells, or manipulating notebook content.

日本語の概要は準備中です。原文の説明を表示しています。

Reference when designing autonomous agent systems or deciding agent freedom/safety/context management/collaboration infrastructure. 10 design principles distilled from Karpathy autoresearch (autonomous ML experiment loop) + AgentHub (agent-native Git DAG). Triggered by "autoresearch", "agenthub", "ratchet" ("래칫"), "context budget", "dumb platform", "program.md", "overnight", "autonomous experiment" ("자율 실험"), "branchless", "single-file constraint" ("단일 파일 제약"), "fixed time budget" ("고정 시간 예산") keywords.

日本語の概要は準備中です。原文の説明を表示しています。

Source-grounded catalog of dynamic terminal UI/UX patterns in Claude Code 2.1.170 and Codex CLI 0.142.4 — live status lines, spinner verb sets, streaming and progress affordances. Reference for GEODE CLI/REPL UX work.

日本語の概要は準備中です。原文の説明を表示しています。

Create and maintain agent skills with concise trigger metadata, judgement-aware guidance, progressive disclosure, and tested interfaces. Use when adding, refactoring, reviewing, or troubleshooting skills and their bundled scripts or references.

日本語の概要は準備中です。原文の説明を表示しています。

Cross-reference diffs to update README, ARCHITECTURE, CHANGELOG.

日本語の概要は準備中です。原文の説明を表示しています。

host / 辅助段优化模块（L3）：把**非 DiT、非解码段**的固定开销落成可复验的优化—— ① 交付与搬运（mp4/图片编码、worker→API 通路、落盘、异步化）； ② 装载与预热（权重加载、编译与图下发预热、镜像/容器预热）。 当用户说“mp4 编码/传输/落盘慢”“交付这段占了不少时间”“模型加载慢”“首启编译太久” “固定开销占比高”“每个请求都要重新编码/重新写盘”时使用。 **入口条件**：占比已达门限（瓶颈标签 `非DiT-host段`；标签枚举、门限与测点口径的单一真源在 model-auto-optimization/references/bottleneck-labels.md，本技能只消费不定义）。 near-miss（看似相关但不属本技能）： - **并发 / 吞吐 / 批处理调度 / 队列深度与限流** → **不属本技能**（会改变所有性能口径与 SLO 语义， 未来独立立项；本技能只在**不改变请求并发模型**的前提下压固定开销）； - VAE/TAE 解码段（计算或通信）→ `vae-opt`；DiT 计算 → `dit-perf-opt`； 多卡并行 / 通信掩盖 → `dit-parallel-opt`； - **装库 / 下载权重 / 构建镜像 / 创建容器** → `env-install`（本技能只优化装载与预热**过程本身**）； - 编译与图下发的**机制本体**（pattern / Inductor / NPUGraph / 图 pool / 缓存条目）→ `pattern-dev` / `aclgraph-dev`（本技能不另做一套缓存）； - 产物一致性与精度判据 → `accuracy-gate`；收益口径 → `perf-gate`； - 瓶颈未定位 / 没有占比锚点 → `model-auto-optimization`。

日本語の概要は準備中です。原文の説明を表示しています。

DiT 计算模块（L3）：把**已定位的 DiT 计算瓶颈**落成特性级选档与实施—— 量化档（W8A16 / W4A16 / W8A8 系列 / W4A4 / MXFP8 / FA 量化）、稀疏（rf_v2 / ada_bsa）、 缓存（DiTCache / AttentionCache / 时间步优化）、编译启用（MindieSDBackend / Pattern 融合 / ACLGraph） 的**开不开、开哪一档、怎么开、怎么复验**；依据是 `docs/zh/features/*`（特性真源）+ framework-integration/references/framework-support-matrix.md（支持状态）。 即使用户只说"这个模型怎么加速""量化/稀疏/Cache 怎么选怎么开""要不要开量化、开哪一档""这个档位开了有没有效果" 而未提 profiling，也应触发。 **入口条件**：瓶颈点已明确（用户带一句实测锚点，或编排层交付标签）时由域入口 `performance-optimization` 按标签分发到本技能；**瓶颈未明（"怎么加速 / 跑通 / 采 profile"）先走 `model-auto-optimization` 定位**，不在本技能内做占比分析。 near-miss：多卡并行形态 / 通信掩盖 / TP·offload 选型 → `dit-parallel-opt`；VAE 解码段与 host 固定开销 → 各自模块（VAE / host）；单算子实现级实测选型（mindie_bench）→ `benchmark-dev`； 需要新增 pattern / 算子才能落地本档 → `pattern-dev` / `operator-dev`；框架侧开关与使能验证 → `framework-integration`；量化器位级契约与精度对齐（编码公式 / 舍入 / scale 粒度）→ `quantization-dev`；精度验收判据 → `accuracy-gate`；数字入库口径 → `perf-gate`。

日本語の概要は準備中です。原文の説明を表示しています。

分布式并行策略选型与实测（USP / CP 通信掩盖 / CFG / TP/RSP/PP 概览；含拓扑相关选型 （按实测拓扑分域条件化：域内 bulk vs 跨域 head-parallel 翻转）与 AlltoAllV 缺陷绕过）。在 model-auto-optimization 中承担 S3：优先 USP、结合拓扑带宽差异选 CP，少量 step + 多 rank 验证特性开启与掩盖（产物：并行方案 + 多 rank 证据）。当用户需要多卡并行 策略选择、**序列并行形态抉择**（纯 Ulysses vs 复合 AllGather-KV×Ulysses：按 GQA / 跨域带宽 / 形态 plumbing 条件化定胜负）、**并行 × 稀疏叠加**（seam 契约：先汇聚后稀疏、 窗口偏移、块对齐、per-head 掩码；含「稀疏看似生效实则未生效」判定）、通信掩盖调优 （含**掩盖率上限**：1-1/n 何时成立、c/f 决定的真实上限、没生效的排查）、**并行方案 差异归因**（阶段 Δ 分解 / 集合通信按 communicator 归属 / 4→8 卡线性度），或排查多卡 跑不动 / 通信暴露大 / 换卡组 / 端口 bind / HCCL 带宽验证问题时使用；**并收编原并行作用域诊断**： 改了 SP/CP/Ulysses/AllGather-KV 后**不报错但结果没变/性能没变**、或小规模能跑大规模崩 （如 Ascend EE1003 coreDim 超限）时，用本技能证明"改动到底有没有生效"（判别量逐层收窄 + 两侧对照 + 日志≠生效，见 `references/scope-effectiveness-check.md`）； 即使用户只说 "多卡跑不动""通信暴露大""为什么没达到 6/7 的掩盖率""CP 和 USP 该选哪个""CP 叠稀疏 怎么不生效"而未说并行，也应触发。特性档位/接口事实见 `docs/zh/features/parallelism.md` /`usp.md`（仓内真源），框架侧开启见 framework-integration； 本技能承载选型决策、monkey-patch 掩盖与多卡诊断实测。 由 dev-workflow 多卡场景触发，亦由 model-auto-optimization 的 S3 阶段触发。

日本語の概要は準備中です。原文の説明を表示しています。

**精度验收标准**：凡是"改动不应改变结果"的场合（等价替换、算子/子模块融合、并行切分、 编译与图下发、拷贝消减），都按本标准判"合格 / 不合格"——等价分层（L1 逐位 / L2 数值门 / L3 有损门）+ 三级验收序（① 同配置重跑逐位 → ② 跨配置数值门 + 产物 md5 不变 → ③ 质量门）+ 判据不达标时的排障入口。当用户说"这个算子能不能换个写法/换个 kernel/等价实现/ 无损替换""替换后结果会不会变""怎么证明逐位一致""结果不对""花屏""尾部塌了" "CPU 跑对 NPU 跑不对"，或发现某个 hotspot 占了大头（例如某类算子在阶段里占 90% 以上）想动手时， 都应触发；即使用户只说"这样改有没有把结果改坏""能不能判它是无损的"也应触发。 覆盖：等价分层与判据、满足本标准的实现约定（per-shape 对拍写进实现，把索引/相位/顺序类重写 错误在首次调用抓住）、per-shape 条件性等价（同一改动在不同形状下结论可能不同）、 判据不达标 → `references/silent-failure-localization.md` 排障入口、否决案例的形态学 （换写法未换 kernel / 差异极小仍非逐位 / 等价但 OOM）、以及收益口径的归属（性能数字一律交 `perf-gate`）。 **近义分流**：选量化档 / 特性选档（要不要开量化、开哪一档）→ `dit-perf-opt`； 并行选型（USP / CP / TP 怎么切）→ `dit-parallel-opt`；本技能只判"结果是否被改变"， 不负责选档与选型。

日本語の概要は準備中です。原文の説明を表示しています。

Drafts USMCA/NAFTA Certificates of Origin for preferential tariff treatment on US-Mexico-Canada goods. Verifies origin criteria, HS classifications, and regional value content. Use when preparing trade certificates, customs origin documentation, duty preference claims, or cross-border shipment certifications.

日本語の概要は準備中です。原文の説明を表示しています。

Guides the agent through TRID tolerance compliance under 12 CFR § 1026.19(e)(3), comparing Closing Disclosure fees to Loan Estimate fees across zero, 10% cumulative, and unlimited tolerance categories. Use when reviewing a CD for tolerance variances, determining whether a revised LE is permitted, calculating cure amounts, or resolving post-closing tolerance violations.

日本語の概要は準備中です。原文の説明を表示しています。

Generates California-compliant privilege logs for withheld or redacted discovery materials under CCP 2031.240. Use when the user mentions privilege log drafting, California privilege log, CCP 2031.240, attorney-client privilege logging, work product designation, Evid. Code 952/954, CCP 2018.030, redaction logging, meet-and-confer privilege disputes, waiver analysis, common interest doctrine, or asks for help documenting withheld documents. Also trigger on references to Hernandez v. Superior Court, Costco Wholesale, or Wellpoint.

日本語の概要は準備中です。原文の説明を表示しています。

Drafts a verified U.S. appellate Bill of Costs itemizing taxable costs under FRAP 39 or state analogs. Trigger when the user needs to prepare, file, or tax appellate costs after judgment, or references FRAP 39, bill of costs on appeal, or appellate cost taxation.

日本語の概要は準備中です。原文の説明を表示しています。

Drafts a mathematically balanced ALTA Settlement Statement for U.S. real estate closings, allocating debits and credits between buyer and seller with prorations, payoffs, lender fees, title charges, and transfer taxes. Enforces balance verification (seller credits minus debits equals cash to seller; buyer debits minus credits equals cash from buyer) and cross-references title commitment, loan documents, and purchase agreement. Use when preparing closing financials, settlement statements, HUD-equivalent documents, or real estate closing allocations.

日本語の概要は準備中です。原文の説明を表示しています。

Generates a cross-referenced U.S. nonprofit annual filing package for state charity-bureau registration. Produces Full Compliance Package, Form-Field Mode, or Executive Briefing with source citations, disclosure cross-checks, and filing instructions. Trigger when the user drafts annual report submissions, renewal packets, penalty-of-perjury certifications, or charity bureau filings. Also trigger on Form RRF-1, CHAR500, BCO-10, nonprofit annual report, charity registration renewal, or fundraiser disclosure attachments.

日本語の概要は準備中です。原文の説明を表示しています。