Show how much context window context-mode saved this session. Displays token consumption, context savings ratio, and per-tool breakdown. Read-only — shows stats only, no reset capability. To wipe the knowledge base entirely, use ctx_purge instead. Trigger: /context-mode:ctx-stats
Guide for working with Feast (Feature Store) — defining features, configuring feature_store.yaml, retrieving features online/offline, using the CLI, and building RAG retrieval pipelines. Use when the user asks about creating entities, feature views, on-demand feature views, stream feature views, feature services, data sources, feature_store.yaml configuration, feast apply/materialize commands, online or historical feature retrieval, or vector-based document retrieval with Feast.
Ilya Sutskever的思维框架与表达方式。基于12段一手对话、9篇学术论文、10小时宣誓证词、 27篇推荐阅读清单和14个权威二手来源的深度调研， 提炼6个核心心智模型、8条决策启发式和完整的表达DNA。 用途：作为思维顾问，用Ilya的视角分析AI技术方向、安全策略、研究品味。 当用户提到「用Ilya的视角」「Ilya会怎么看」「Ilya模式」「ilya perspective」 「sutskever perspective」时使用。 即使用户只是说「帮我用Ilya的角度想想」「如果Ilya会怎么做」「切换到Ilya」也应触发。
塔勒布(Nassim Nicholas Taleb)的思维框架与表达方式。基于40+个来源的深度调研， 提炼6个核心心智模型、9条决策启发式和完整的表达DNA。 用途：作为思维顾问，用塔勒布的视角分析问题、审视决策、质疑主流叙事。 当用户提到「用塔勒布的视角」「塔勒布会怎么看」「塔勒布模式」「反脆弱视角」「taleb perspective」时使用。 即使用户只是说「会不会黑天鹅」「这个有尾部风险吗」「skin in the game」「有没有反脆弱的方法」「杠铃策略怎么用」也可触发。 不要在用户只是做一般风险评估或问「靠不靠谱」时触发——只在涉及极端风险、反脆弱、预防原则等塔勒布核心概念时激活。
NVIDIA's runtime safety framework for LLM applications. Features jailbreak detection, input/output validation, fact-checking, hallucination detection, PII filtering, toxicity detection. Uses Colang 2.0 DSL for programmable rails. Production-ready, runs on T4 GPU.
Evaluates code generation models across HumanEval, MBPP, MultiPL-E, and 15+ benchmarks with pass@k metrics. Use when benchmarking code models, comparing coding abilities, testing multi-language support, or measuring code generation quality. Industry standard from BigCode Project used by HuggingFace leaderboards.
Evaluates LLMs across 60+ academic benchmarks (MMLU, HumanEval, GSM8K, TruthfulQA, HellaSwag). Use when benchmarking model quality, comparing models, reporting academic results, or tracking training progress. Industry standard used by EleutherAI, HuggingFace, and major labs. Supports HuggingFace, vLLM, APIs.
Visualize training metrics, debug models with histograms, compare experiments, visualize model graphs, and profile performance with TensorBoard - Google's ML visualization toolkit
Facebook's library for efficient similarity search and clustering of dense vectors. Supports billions of vectors, GPU acceleration, and various index types (Flat, IVF, HNSW). Use for fast k-NN search, large-scale vector retrieval, or when you need pure similarity search without metadata. Best for high-performance applications.
Extract structured data from LLM responses with Pydantic validation, retry failed extractions automatically, parse complex JSON with type safety, and stream partial results with Instructor - battle-tested structured output library
LLM observability platform for tracing, evaluation, and monitoring. Use when debugging LLM applications, evaluating model outputs against datasets, monitoring production systems, or building systematic testing pipelines for AI applications.
Open-source AI observability platform for LLM tracing, evaluation, and monitoring. Use when debugging LLM applications with detailed traces, running evaluations on datasets, or monitoring production AI systems with real-time insights.
Check CI status, analyze failures, and explain skips for a Dynamo PR
Best practices for using the stagewise built-in JavaScript sandbox. Explains how to access APIs for browser debugging/interaction, use external dependencies, file system access, running mini-apps, etc.
Check and optimize MetaDescription frontmatter fields in VS Code documentation. Use when auditing, adding, or improving page descriptions for SEO and discoverability. Apply this when making content changes to markdown articles.
Resolve inline code review comments on a GitHub PR. Use when asked to "resolve review comments", "address PR feedback", "fix PR comments", or "work through review comments". Fetches every inline comment with its surrounding code context, then applies each change systematically.
Find all FIXME comments across the codebase and attempt to resolve them. Use when the user asks to fix, resolve, or address FIXME comments, or when running the "fixme" command. Runs a script to locate every FIXME with surrounding context (2 lines before, 5 lines after) and then works through each one systematically.
Query patterns including pagination, search, filtering, and specifications for FSH. Use when implementing GET endpoints that return lists or need filtering.
Add, modify, or debug benchmark scenarios for measuring Materialize performance. Covers three frameworks: Feature Benchmark (single-operation micro-benchmarks), Scalability Test (SQL throughput under concurrency), and Parallel Benchmark (sustained latency over time via scenarios.py). Trigger on "benchmark", "feature benchmark", "scalability test", "parallel benchmark", "performance regression", "micro-benchmark", "TPS", "latency test", or when editing files in feature_benchmark/scenarios/, scalability/workload/workloads/, or parallel_benchmark/scenarios.py. Note: this is about benchmark measurement frameworks, not the parallel-workload stress-testing framework (which tests for panics under concurrency, not performance).
Investigate CI failures on a PR using gh and bk CLI tools. Trigger when the user asks about failing checks, Buildkite failures, or CI issues — including casual phrases like "why is CI red", "build broken", "checks failing", "what went wrong in CI", "nightly broke", "tests failing on this PR", or pastes a Buildkite URL. Also trigger when the user mentions a specific PR number and wants to understand why it's failing.
Extend the parallel-workload stress-testing framework, which runs random SQL actions concurrently to catch panics and unexpected errors (not performance measurement — see mz-benchmark for that). Trigger when the user mentions "parallel workload", "parallel-workload", "action.py" in the context of parallel workload, or wants to test for panics or unexpected query errors under concurrency. Use this skill even if the user just says "add this to parallel workload" or references a bug that panics under concurrent DDL/DML.
