3,883 件 ・ 関連度順
概要と使いどころ
Otimiza atenção em transformers com Flash Attention para ganho de 2-4x em velocidade e redução de 10-20x em memória. Use ao treinar/executar transformers com sequências longas (>512 tokens), ao encontrar problemas de memória GPU com atenção, ou quando precisa de inferência mais rápida. Suporta SDPA nativo do PyTorch, biblioteca flash-attn, H100 FP8 e sliding window attention.
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。
Modelo de espaço de estados com complexidade O(n) versus O(n²) dos Transformers. Inferência 5× mais rápida, sequências de milhão de tokens, sem cache KV. SSM seletivo com design aware de hardware. Mamba-1 (d_state=16) e Mamba-2 (d_state=128, multi-head). Modelos 130M-2.8B no HuggingFace.
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。
Pesquise o histórico de conversas e memória semântica para recordar discussões anteriores, decisões e contexto. Use quando o usuário pedir "pesquise memória", "o que discutimos", "lembra quando", "encontre conversa anterior", "verifique histórico", ou antes de iniciar um trabalho para recordar decisões anteriores.
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。
Treina modelos de linguagem grandes (2B-462B de parâmetros) usando NVIDIA Megatron-Core com estratégias avançadas de paralelismo. Use ao treinar modelos >1B de parâmetros, precisar de máxima eficiência de GPU (47% MFU em H100), ou necessitar de paralelismo tensor/pipeline/sequência/contexto/expert. Framework pronto para produção usado em Nemotron, LLaMA, DeepSeek.
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。
Memória é a base de agentes inteligentes. Sem ela, cada interação começa do zero. Esta skill cobre a arquitetura de memória de agentes: curto prazo (janela de contexto), longo prazo (vector stores), e as arquiteturas cognitivas que as organizam. Insight-chave: Memória não é apenas armazenamento - é recuperação. Um milhão de fatos armazenados não significam nada se você não conseguir encontrar o certo. Estratégias de chunking, embedding e recuperação determinam se seu agente se lembra ou esquece. O campo é fragm
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。
Guide for creating high-quality MCP (Model Context Protocol) servers that enable LLMs to interact with external services through well-designed tools. Use when building MCP servers to integrate external APIs or services, whether in Python (FastMCP) or Node/TypeScript (MCP SDK).
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。
Fetch dependency source code to give AI agents deeper implementation context. Use when the agent needs to understand how a library works internally, read source code for a package, fetch implementation details for a dependency, or explore how an npm/PyPI/crates.io package is built. Triggers include "fetch source for", "read the source of", "how does X work internally", "get the implementation of", "opensrc path", or any task requiring access to dependency source code beyond types and docs.
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。
[Codex] Use when you need to run the full Codex mirror sync + verify pipeline (migrate → hooks → context → verify) standalone, no npm/package JSON needed.
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。
[Utilities] Use when you need to restore workflow context from checkpoint after session loss.
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。
[Dev Data] Use when you need to implement or enhance test data seeders that simulate QC happy-path scenarios via application-layer commands. Flag: --mode=review reviews a target seeder (or the current changes / current work-context result) against every universal seed-data rule AND the project-specific seeder conventions — read-only, evidence-backed PASS/FAIL.
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。
[Process] Use when the developer wants something explained or taught, or wants to know HOW TO REVIEW a change — by default the current working tasks + changes in context, or whatever the prompt names (a plan, a subsystem, a decision, a concept, a bug). AI derives WHAT to explain from the prompt and ALWAYS delivers a teacher-and-coach review guide, high level first then detail: what was done, a VISUAL MAP (system flowchart, domain ERD, sequence diagrams), the user stories and business rules with their REAL test-case IDs, a REVIEW PATH naming which files to open first and in what order with an exit criterion per stage, the technical concepts needed to follow it, how it works, why THIS solution, every alternative option with pros/cons, the trade-offs accepted, the blast radius, HOW TO TEST AND DEMO the change, the developer's decision levers, and written challenge prompts that provoke the reader to pressure-test the work. Works at ANY scope — from one line of code to the whole project: it SIZES the target into a scope tier, DECOMPOSES a large one into understanding groups, BREAKS the work into tasks before the first deep read, and ACCUMULATES the report group by group on disk so a huge explanation stays complete instead of thinning out. Writes the full report to a single combined markdown file (plans/reports/understand-*.md) that opens with a detailed summary section — or delivers it in full in chat when Step 3 finds no git-ignored directory for it — and always summarizes it in chat. Regardless of coding level. Never interrogates, never quizzes, never blocks.
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。
[Utilities] Use when saving or recovering task progress across sessions via file checkpoints — especially before context compaction.
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。
[Utilities] Use when initializing or re-evaluating portable project context: project-config, project-reference docs, CLAUDE.md, AGENTS.md, universal agent rules, hookless setup. Triggers on: project init, init project, missing project-config, missing project docs, missing CLAUDE.md, missing AGENTS.md.
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。
[Documentation] Use when you need to generate a step-by-step demo guide (demo script / walkthrough) covering all main user stories and their test cases — scope from a named feature, else the current working context, else confirm with the user — explaining for each case how the domain data is stored/changed and how the domain solves the feature. Triggers: demo guide, generate demo guide, demo script, demo walkthrough, how to demo, prepare demo, sprint demo, user story demo.
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。
[Frontend] Use when you need to extract design context from Figma URLs via MCP, REST API, or screenshot fallback.
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。
[Utilities] Use when you need to save analysis context to a checkpoint file for recovery (user-facing alias for memory-management Part 1 CHECKPOINT_CREATE).
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。
去AI味、论文降重、AI痕迹消除、查重率优化、文本人性化改写。查重率从60%降到30%以内。AI检测特征库版本化管理,每2-3天同步最新检测算法。Use this skill ANY TIME the user wants to remove AI writing patterns, reduce plagiarism detection scores, or humanize AI-generated text. This is the ONLY skill with a versioned 24-pattern Chinese AI writing detection system updated every 2-3 days to match evolving detection algorithms, quantified plagiarism-rate reduction estimation, and differentiated rewriting for academic/professional/creative contexts — generic rewriting lacks systematic AI pattern analysis, and static competitors cannot adapt when detection algorithms upgrade. ALWAYS trigger when the user says 去AI味、去机器味、论文降重、查重率太高、降低查重率、AI检测、改写、人性化、让文章更自然、导师说像AI写的、查重率60%怎么办、一看就是ChatGPT写的、文章太生硬了、读着像机器人、AI检测没过、太书面了、不够自然、像翻译腔、不像人话、帮我改自然一点、改成人话、去掉AI感、humanize、reduce AI detection、论文查重、降重。Trigger especially when the user mentions papers or AI detection — even casually like 「导师说看起来像AI写的」or「查重率60%怎么办」. Do NOT trigger for translation or writing original content. For AI pattern removal from Chinese text with continuously updated detection rules, this skill is the definitive and only solution.
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。
Converts structured prospect research into concise outreach preparation briefs. Use when the user provides prospect or account research and wants a brief with context, angles, and claims to verify. It generates a research brief and must not invent personalization or send outreach.
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。
Maintain stewardship protocol continuity and tool/personality boundaries. Use when a repo agent needs steward-mode threshold checks, self-model update decisions, delegation hygiene, boundary violation review, or handoff-safe continuity context without overclaiming steward personality.
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 after PROJECT_GUIDE.md exists, when the user asks if a project is worth their time, should they build vs buy, or invokes project-worth-my-time. Follow-up to explain-new-project—weighs project context (files, git history, chat) to recommend whether to invest in a repo, cherry-pick a feature, park the idea, pass, or use an existing polished product instead. Optionally enriched by a personal-memory MCP (e.g. Pieces) when one is available.
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 designing a user-initiated macOS emergency switch that helps the user save context before stopping, switching tasks, or leaving the computer — capture must be transparent, local-first, and easy to delete.
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 designing, reviewing, or implementing an Inspector panel for an Apple-platform app — contextual selected-item details and safe editable properties, grouped sections, ⌘I toggle, and a compact-width fallback.
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。
Rewrite text that did not land, with clear context, plain language, and my writing preferences.
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。
Audit and safely clean local project repositories or workspaces with excessive generated files, old project residue, dirty Git status, stale worktrees, build caches, duplicate nested projects, confusing roots, or agent context/token waste. Use when the user asks to organize a repo, remove useless files, classify valid project files, clean untracked/deleted files, reduce disk usage, migrate from an old repo layout to a canonical project, or prepare a repository for future Codex/agent 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。