Skill
agent-skills
Gonglitian/agent-skills
46
InicialPluginScore · de 0 a 100A collection of agent skills for Claude Code, Cursor, and other AI coding tools
Instalação
$
npx convez-ai Gonglitian/agent-skills1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
Estrelas
2Forks
0Acompanhando
0Saúde
28%Último commit
há 4 mesesVersão
—Issues abertas
0Contribuidores
2README
A tradução para o português deste README ainda está na fila.
Agent Skills Collection
A comprehensive collection of 45 Claude Code skills for ML research workflows, covering the full research cycle from literature survey to experiment management.
Installation
# Install all skills
npx skills add Gonglitian/agent-skills
# Install specific skill
npx skills add Gonglitian/agent-skills --skill read-paper
Skills Overview
Research & Literature (研究调研)
| Skill | Description |
|---|---|
| comprehensive-survey | Full-spectrum survey: 6 academic sources + 5 social platforms → parallel subagents → structured report with concept glossary |
| topic_survey | Interactive field survey: explore sub-topics guided by user, produce a structured literature review / reading list |
| read-paper | Deep-read papers with VLM figure analysis, parallel subagents, and structured 12-section notes |
| gap-to-method | Multi-dimensional literature matrix → Gap discovery → Evidence-based method proposal |
| idea_refinery | Iteratively refine research ideas through survey, validation, and branching exploration |
| paper_related_works | Build predecessor/successor citation maps for any paper |
| create_skill_with_paper | Turn academic papers into reusable skill reference cards |
| paper-intro-writing | Write reviewer-grade paper Introductions: 6-段黄金骨架 + 10 篇 VLA/robot-learning 顶会 paper 拆解 + Mad-Libs 填空模板 + 25 项自检表 |
| research-refine | Turn a vague direction into a problem-anchored, elegant method plan via iterative GPT-5.4 review |
| experiment-plan | Turn a refined proposal into a claim-driven experiment roadmap: ablation matrix, eval protocol, run order, compute budget |
| review-ral | IEEE RA-L paper review assistant: read PDF → multi-source related-work search → parallel deep-read → full bilingual review with scores |
| multi-platform-search | Cross-platform info gathering from Xiaohongshu, Bilibili, Zhihu, and X/Twitter simultaneously |
| notion-paper-table | Build a structured literature survey database in Notion with paper metadata |
| paper-discovery-sources | Shared 3-source paper-discovery reference (vec-db / Semantic Scholar / AlphaXiv) loaded by other skills — not user-invokable |
| arxiv-deepdive | Code-grounded paper deep-dive: arXiv PDF + official repo shallow-clone (fetch.sh, size-capped) → 01_highlevel (core idea) + 02_technical (training/inference details pinned to file:line, runnable commands, repro pitfalls); flags paper-vs-code mismatches, batch fan-out with INDEX.md |
Paper Knowledge Base (论文知识库)
| Skill | Description |
|---|---|
| omnibox-search | Semantic search over 965 paper reports + 353 video transcripts: full-report chunk index (29k+, bge-m3 + sqlite-vec) → top-k with hit section + note path, topic/受控 tag filters (expects local OmniBox KB at ~/proj/omnibox) |
| omnibox-sync | Token-minimal incremental sync of an OmniBox (小黑) public share: metadata-only diff (id + updated_at) vs persistent snapshot → fetch content for delta only → JSON + MD delta files (new / changed / removed) — ingestion stage feeding arxiv-deepdive |
| omnibox-video | Video 预阅读 pipeline: OmniBox share 里的 B站/小红书 videos → audio download (yt-dlp + XHS-Downloader signed API) → faster-whisper GPU transcription → mine spoken arXiv papers (ASR name correction) → dedupe vs existing KB → arxiv-deepdive reports + INDEX; chains omnibox-sync + audio-transcribe + arxiv-deepdive |
Experiment Management (实验管理)
| Skill | Description |
|---|---|
| auto_experiment | Full experiment lifecycle: workspace setup → iterative experiment loop → final report |
| experiment_report | Structured experiment reports with results-first format and W&B integration |
| gpu-train-monitor | Multi-GPU training monitoring: GPU stats, loss tracking, throughput analysis |
| train-debug | Systematic training diagnosis: OOM, NaN gradients, loss plateau, multi-GPU issues |
| data-pipeline-check | Dataset validation, schema checks, quality metrics, and compatibility verification |
| weights-and-biases | Track ML experiments with W&B: auto-logging, real-time visualization, hyperparameter sweeps, model registry |
| evidence-report | KDA measurement layer: Measure → Diagnose → Plan, never guess — immutable profile/<run>/ dirs + aggregate.py (mean±stddev, baseline delta) + 5 analysis dimensions + signal→cause→fix playbook → 2–4 recommendations ranked by evidence × impact |
Engineering (代码工程)
| Skill | Description |
|---|---|
| project-init | Initialize research projects: conda env, CLAUDE.md, git, data path organization |
| setup-dev-env-ubuntu | Bootstrap a fresh Ubuntu machine: zsh + oh-my-zsh + fzf, Ghostty, Edge, VSCode, Miniconda, Claude Code + claude-hud, gh auth, Snipaste, WeChat, fcitx, Tailscale |
| tmux-workspace | Generate tmuxinator configs for multi-project terminal workspaces |
| ucr_hpcc_cluster | Work with the UCR HPCC cluster: connecting, job submission, software management, data storage |
| ghostty-cjk-input-debug | Diagnose and fix CJK input-method issues in Ghostty terminal on Linux (snap + fcitx5 + GTK4) |
| kda | KDA (Kernel Design Agents) orchestration entry point: engineering-loop (control) + domain-wiki (knowledge) + evidence-report (measurement) → Task Contract → correctness-first → evidence-guided optimization → per-regime specialization → promotion, with anti-reward-hacking guards |
| engineering-loop | KDA-style long-horizon optimization harness: task contract → draft/plan → one-candidate-at-a-time iterate → promote; immutable baseline + candidates.jsonl DAG (helper script) + benchmark.csv + 3-phase search + anti-reward-hacking guards |
| domain-wiki | Provenance-tracked domain knowledge base: sources → atomic entries (verbatim quote + source_ref + date + confidence) → 4-axis query (tag/type/source/symptom) + staleness check via 2 stdlib-only scripts; knowledge layer of the KDA loop |
Meta (元技能)
| Skill | Description |
|---|---|
| skill-creator | Create new skills, run evals, benchmark variance, optimize descriptions for triggering accuracy |
| find-skills | Discover and install skills from the open agent-skills ecosystem (skills.sh) |
| planning-with-files | Manus-style file-based planning (task_plan.md / findings.md / progress.md) for complex multi-step tasks |
| long-horizon-spec | 协作式 Plan Mode for long tasks: refine intent → read-only exploration → multi-round AskUserQuestion interview → runnable acceptance checks → SPEC.md hard gate (user must approve) → pick autonomy engine (/goal, Stop hook, Dynamic Workflow, or engineering-loop handoff) + adversarial final review |
Domain-Specific (领域专用)
| Skill | Description |
|---|---|
| isaaclab-dev | Isaac Lab robot simulation development reference manual |
| isaaclab-async-pipeline-dev | Async data generation pipeline for Isaac Lab |
| wuyin-gpt-image-2 | GPT-Image-2 text-to-image / image edit via 速创API (api.wuyinkeji.com): async submit → poll → download, with 3 paper-figure style presets (UniVLA / Physical Intelligence / Fast-WAM), each shipping curated reference images auto-uploaded to catbox and reused as urls for high-fidelity style-mimic |
| marp-slide | Create professional Marp presentation slides with 7 themes, custom layouts, and auto quality improvements |
| audio-transcribe | Transcribe audio/video to text: faster-whisper large-v3 with parallel CUDA scheduling + mlx-whisper counterpart for Apple Silicon (Metal), plus a context-aware LLM polish pass (auto-triggered, timestamp-preserving) |
Obsidian (笔记系统)
| Skill | Description |
|---|---|
| obsidian-markdown | Obsidian Flavored Markdown authoring: wikilinks + embeds (notes/images/PDF/audio/search) + 13 callout types + YAML properties + tags/comments/Mermaid/LaTeX, with 3 syntax reference sheets |
| obsidian-bases | Author Obsidian Bases (.base YAML): 4 view types (table/cards/list/map) + filters/formulas/groupBy/summaries, with Duration-math gotchas, YAML quoting rules, and a full functions reference (Date/String/Number/List/File/RegExp) |
| obsidian-cli | Drive a live Obsidian vault via the obsidian CLI: notes / search / daily / tasks / properties / tags / backlinks + plugin-dev loop (reload → dev:errors → screenshot/DOM → console) with in-app JS eval and mobile emulation |
Research Workflow
These skills support the full research cycle:
idea_refinery → gap-to-method → project-init → auto_experiment → experiment_report
↑ ↑ ↑
comprehensive-survey read-paper gpu-train-monitor
train-debug
data-pipeline-check
And the local paper knowledge-base pipeline:
omnibox-sync ─→ arxiv-deepdive ─→ omnibox-search
omnibox-video ↗ (01_highlevel + 02_technical) (bge-m3 + sqlite-vec semantic search)
License
MIT