Total 50,653 skills, AI & Machine Learning has 8491 skills
Showing 12 of 8491 skills
Fact-forcing gate that blocks Edit/Write/Bash (including MultiEdit) and demands concrete investigation (importers, data schemas, user instruction) before allowing the action. Measurably improves output quality by +2.25 points vs ungated agents.
W&B integration. Manage data, records, and automate workflows. Use when the user wants to interact with W&B data.
Use when an agent needs to interact with PolyBaskets prediction market baskets on Vara Network — create baskets, place bets, query state, claim payouts, or understand the protocol. Do not use for building Sails programs or general Vara development (use vara-skills for that).
Use this skill whenever deciding what features to extract from raw marketplace assets — listing photos, owner-entered listing metadata, sitter wizard responses — to power item-to-item (similar listings), user-to-item (homefeed ranking), or user-to-user (mutual-fit matching) recommenders in a two-sided trust marketplace. Covers asset auditing, first-principles feature decomposition from the decision the user is making, vision-feature extraction (CLIP, room-type classification, amenity detection, aesthetic and quality scoring), listing text and metadata encoding (categoricals, multi-hot amenities, H3 geo-hashing, sentence-transformer description embeddings, structured pet triples), sitter wizard design (information-gain ordering, multiple-choice over free text, genuine skippability, hard constraint versus soft preference), derived-composition patterns for i2i / u2i / u2u (precomputed ANN shelves, multi-modal fusion, two-tower affinity, symmetric mutual-fit scoring, interpretable subscores), feature quality governance (single registry, training-serving parity, coverage and drift alarms, PII scrubbing, schema versioning), and incremental value proof (one feature at a time, ablation A/B, kill reviews, exploration slice, permanent feature-free baseline). Trigger even when the user does not explicitly say "feature engineering" but is asking how to get more signal out of listing photos, listing metadata, or the sitter onboarding wizard, or how to improve i2i / u2i / u2u quality without blindly ingesting a new model.
Use this skill when the user asks to "evaluate MCP tools", "test tool selection", "improve tool descriptions", "check MCP schema quality", "eval my MCP server", or wants to measure whether Claude uses their MCP tools correctly. Tests tool selection accuracy, analyzes schema quality, and iteratively optimizes descriptions. Companion to build-mcp-server.
MindOS is the user's local knowledge assistant and shared knowledge base. It keeps decisions, meeting notes, SOPs, debugging lessons, architecture choices, research findings, and preferences available across sessions and agents. 更新笔记, 搜索知识库, 整理文件, 执行SOP/工作流, 复盘, 追加CSV, 跨Agent交接, 路由非结构化输入到对应文件, 提炼经验, 同步关联文档. NOT for editing app source, project docs, or paths outside the KB. Core concepts: Space, Instruction (INSTRUCTION.md), Skill (SKILL.md); notes can embody both. Trigger on: save or record anything, search for prior notes or context, update or edit a file, organize notes, run a workflow or SOP, capture decisions, append rows to a table or CSV, hand off context to another agent, check if something was discussed before, look up a past decision, distill lessons learned, prepare context for a meeting, quick-capture to staging area, organize inbox, check knowledge health, detect conflicts or contradictions, find stale content. Chinese triggers: 帮我记下来, 搜一下笔记, 更新知识库, 整理文件, 复盘, 提炼经验, 保存, 记录, 交接, 查一下之前的, 有没有相关笔记, 把这个存起来, 放到暂存台, 整理暂存台, 知识健康检查, 检测知识冲突. Proactive behavior — do not wait for the user to mention MindOS: (1) When user's question implies stored context may exist (past decisions, previous discussions, meeting records) → search MindOS first, even if they don't explicitly mention it. (2) After completing valuable work (bug fixed, decision made, lesson learned, architecture chosen, meeting summarized) → offer to save it to MindOS for future reference. (3) After a long or multi-topic conversation → suggest persisting key decisions and context.
Atlas Cloud API integration skill — quickly call 300+ AI image generation, video generation, and LLM models through a unified API. Use this skill when the user needs to integrate AI image generation (e.g., Flux, Seedream, DALL-E), AI video generation (e.g., Kling, Sora, Seedance), or call LLM APIs (OpenAI-compatible format) into their project. Applicable scenarios include: generating images, generating videos, calling large language models, using Atlas Cloud API, configuring ATLASCLOUD_API_KEY, querying available model lists, searching models by keyword, uploading local images/media files, one-step quick generation, image-to-video, text-to-image, text-to-video, AI content creation tool integration. Even if the user doesn't explicitly mention Atlas Cloud, this skill should be considered whenever AI media generation API integration development is involved.
Token-efficient persistent memory system for Claude Code that extends your session limits by 3-5x. Layered architecture with progressive loading, compact encoding, branch-aware context, smart compression, session diffing, conflict detection, session continuation protocol, and recovery mode. Activates at session start (if MEMORY.md exists), on "remember this", "pick up where we left off", "what were we doing", "wrap up", "save progress", "don't forget", "switch context", "hand off", "memory health", "save state", "continue where I left off", "context budget", "how much context left", or any session start on a project with existing memory files. This skill solves two problems at once: Claude forgetting everything between sessions, AND sessions hitting context limits too fast. It replaces thousands of wasted re-explanation tokens with a compact, structured memory load that gives Claude full project context in under 2,000 tokens.
Apply dual-process theory to diagnose whether judgments arise from fast intuitive (System 1) or slow analytical (System 2) processing and identify resulting cognitive biases. Use this skill when the user needs to explain why quick decisions go wrong, design choice architectures that account for cognitive defaults, audit decision processes for heuristic errors, or when they ask 'why do people misjudge probability', 'how to reduce snap-judgment errors', or 'when does intuition fail'.
Verify and build the required environment for Triton operator development on the Ascend platform, including configurations of dependencies such as CANN, Python/torch/torch_npu/triton-ascend and PATH environment variables. This is used when users need to configure the Triton operator development environment, check the installation of CANN/torch/triton-ascend, or verify whether the environment is available.
Guide Catlass operator performance tuning. Process: Read the Catlass optimization guide, obtain/update profiler baseline, modify tiling according to the guide, recompile, **mandatorily generate and display performance comparison report**, iterate and compare. Tuning strategies are based on Catlass documentation. Ask for clarification if conditions are unclear.
Triage a daily msverl regression run by reading the baseline comparison log, stopping on success, extracting the most relevant training failure evidence from the daily training log when needed, collecting recent commits from verl main and MindSpeed master, and ranking the most likely culprit commits with concise fix-direction guidance.