菜单
Find LLM tokens-per-second (TPS), latency, and throughput metrics. If direct API benchmarking is impossible, pivot to published third-party benchmarks and provide a manual fallback script.
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
Run AI coding agents on disposable repo clones. The agent works on a clone — it can't touch your real repo. An optional container (Podman or Docker) provides build/test isolation. You review the diff and decide what (if anything) to apply.
Create Deezer playlists programmatically from any query — similar artists, genre mixes, festival lineups, mood-based collections. Four-tier data pipeline: Deezer public REST API + Last.fm scrobble data for discovery, GQL Pipe API for smart mixes and playlist creation, web search for subjective curation. Uses ARL cookie auth — no OAuth app required.
Generate compact AI-readable context maps from codebases — tools like codesight, repomix, agentic-context that pre-compute project structure to save tokens in AI coding sessions.
Research-focused query handling with multi-source synthesis, citations, and Obsidian persistence. Like a self-hosted Perplexity/Vane but CLI-native. Best for quick-to-medium lookups using Kagi. Use when the user asks factual questions, needs citations, or wants a direct answer — not a full research report (use deep-research) or social sentiment (use last30days). Triggers on: research, look into, what's the latest on, compare, explain, investigate.
Write articles, guides, blog posts, tutorials, newsletter issues, research reports, and deep research outputs in a distinctive voice derived from supplied examples or brand guidance. Use when the user wants polished written content longer than a paragraph, deep research on a topic, or a research report — especially when voice consistency, structure, and credibility matter. Triggers on: 'write an article', 'research report', 'deep research', 'long-form', 'blog post', 'guide', 'newsletter', 'white paper', 'research paper'.
Systematic research methodology for major consumer durables (appliances, HVAC, power tools, outdoor equipment) and smart garden/outdoor devices (bird feeder cameras, bird baths, smart outdoor gadgets). Emphasis on real reliability data, failure mode analysis, and head-to-head comparison. Use when the user asks to research, review, compare, or evaluate major purchases where longevity and repair risk matter, or when researching smart bird feeders, bird bath cameras, and similar connected outdoor devices. Triggers on: washer/dryer, refrigerator, dishwasher, HVAC, furnace, AC, generator, power tool, appliance reviews, appliance reliability, which [appliance] to buy, compare models, bird feeder camera, bird bath camera, smart garden devices.
| name | llm-performance-benchmark |
| description | Find LLM tokens-per-second (TPS), latency, and throughput metrics. If direct API benchmarking is impossible, pivot to published third-party benchmarks and provide a manual fallback script. |
| trigger | User asks for model speed, TPS, tokens per second, throughput, latency, or performance benchmark of a specific LLM. |
If direct measurement is impossible, immediately search for third-party benchmark data.
| Site | What it tracks | URL pattern |
|---|---|---|
| Artificial Analysis | Provider-level TPS, TTFT, price | artificialanalysis.ai/models/<model>/providers |
| DeepInfra Blog | Provider shootouts (Groq, Fireworks, etc.) | deepinfra.com/blog/<model>-api-benchmarks |
| LLM-Stats | Leaderboard rankings, quality scores | llm-stats.com/models/<model> |
"<model-name>" tokens per second benchmark throughput
"<model-name>" API provider benchmark artificialanalysis
kimi-k2.6 is the current flagship; kimi-k2-0905 / kimi-k2-0711 are earlier K2 series variants. Do not conflate them.Present:
Always give the user a copy-paste script they can run with their own key.