AgentSkillsArxiv

Scaffold and operate an autonomous Triton GPU-kernel optimization loop on FlashInfer-Bench + Modal, modeled on the MLSys-2026 DSA contest winner (Dogacel/auto-gpu-kernel). Provides a ready-to-run template (CLAUDE.md, .claude/commands/{optimize,benchmark,log-experiment}, .claude/agents/{profiler,research,workload-inspector}, Modal benchmark + paired A/B harness, packed-solution scripts), plus customization guidance for retargeting to a different FlashInfer-Bench kernel definition. Use when the user wants to (1) bootstrap an autonomous Triton-kernel optimizer for any FlashInfer-Bench kernel, (2) set up an experiment-logged optimization loop with sub-agents for profiling, workload inspection, and plateau-breaking research, (3) run cloud-only (no local GPU) GPU benchmarking via Modal B200, or (4) reproduce or adapt the auto-gpu-kernel architecture for FlashInfer competition tracks or similar kernel-optimization workflows.

Provides guidance for writing and benchmarking optimized CUDA kernels for NVIDIA GPUs (H100, A100, T4) targeting HuggingFace diffusers and transformers libraries. Supports models like LTX-Video, Stable Diffusion, LLaMA, Mistral, and Qwen. Includes integration with HuggingFace Kernels Hub (get_kernel) for loading pre-compiled kernels. Includes benchmarking scripts to compare kernel performance against baseline implementations. Use for kernel types: attention, rmsnorm, rope, adaln, geglu, benchmark, transformers, diffusers, huggingface-kernels, get_kernel.

Investigate a consistently failing SGLang CI test to find the root cause — whether it is a code regression from a specific PR, a hardware/runner-specific issue, or an environment change. Optionally reproduces the failure on a remote GPU server over SSH + Docker. Use when an SGLang CI test is failing on main, you need to bisect which PR introduced a regression, you suspect a runner-specific or GPU-specific issue, or you want to reproduce a CI failure on a remote server.

Step-by-step tutorial for adding new CUDA kernels to FlashInfer

Write a Statement on Education (Teaching Statement / Teaching Philosophy / Teaching Vision / "onderwijsvisie") for an academic faculty application, especially in Computer Science / AI / Data Science at research universities. Covers structure (1–3 page format), evaluation rubrics, evidence-based pedagogy, supervision frameworks, US vs Dutch/EU framing, TU Delft programme curricula (BSc CSE, MSc CS, MSc DSAIT) with verified course mapping, real-world example anatomies, and a section-by-section drafting workflow that produces a credible, specific, non-generic statement. Use whenever an applicant must write or revise a teaching statement, education statement, teaching vision, teaching philosophy, or onderwijsvisie — particularly for TU Delft, ETH, KTH, Cambridge, Imperial, or any EU/NL CS faculty position.

Generate an Obsidian-format CUDA StudyVault for learning NVIDIA's GPU programming stack. Three modes: (1) Curriculum mode (default, no source files needed) builds a prereq-ordered vault around the 6-topic learning path: CUDA kernels (threads/blocks/warps, memory hierarchy, TMA, WGMMA, cp.async), CUTLASS + CuTe (Layout/Tensor, MMA atoms, GEMM pipelines), cuTile (Python tile DSL), open-gpu-kernel-modules (RM, GSP firmware, UVM, kernel-open), NCCL (collectives, NVLink-SHARP, transports), and NVSHMEM (PGAS, symmetric heap, IBGDA); (2) Codebase mode generates an onboarding vault from a real CUDA/CUTLASS/NCCL/NVSHMEM/driver source tree; (3) Document mode turns NVIDIA whitepapers, GTC slide decks, and PDFs into study notes. Mode is auto-detected from CWD markers. Use when the user wants to learn CUDA from scratch, onboard onto a GPU codebase, build an exercise list for CUDA topics, or preprocess NVIDIA PDFs/docs into structured study material. Pair with the `cuda-tutor` skill to quiz against the resulting vault.

Interactive quiz tutor for a CUDA StudyVault built by `cuda-tutor-setup`. Delivers 4-question rounds with concept-level proficiency tracking (🟥/🟨/🟩/🟦/⬜) across the 6 CUDA learning topics: CUDA kernels (threads/blocks/warps, memory hierarchy, TMA, WGMMA, cp.async), CUTLASS + CuTe (Layout/Stride/Tensor, MMA atoms, GEMM pipelines), cuTile (Python-first tile DSL), open-gpu-kernel-modules (RM, GSP firmware, UVM, kernel-open layout), NCCL (collectives, topology, NVLink-SHARP, transports), and NVSHMEM (PGAS, symmetric heap, IBGDA, on-stream API). Use when the user wants to (1) take a diagnostic CUDA assessment, (2) drill weak GPU concepts, (3) study a specific CUDA topic, (4) review the learning dashboard, or says things like "quiz me on CUDA", "test my CUTLASS knowledge", "drill NCCL", "/cuda-tutor", "퀴즈".

Generate an Obsidian-format Triton StudyVault for learning OpenAI Triton's GPU programming stack. Three modes: (1) Curriculum mode (default, no source files needed) builds a prereq-ordered vault around 6 topics: Triton basics, tiling & autotuning, matmul patterns (tiled GEMM, split-K, FP8 `tl.dot`), attention & reductions (FlashAttention), compiler internals (TTIR/TTGIR/LLIR, NVIDIA/AMD backends, WGMMA + TMA lowering), and ecosystem/production (`torch.compile`/Inductor, AOT, `proton`); (2) Codebase mode generates an onboarding vault from a real Triton/Inductor source tree; (3) Document mode turns Triton docs, blog posts, and tutorial PDFs into study notes. Mode is auto-detected from CWD markers. Use when the user wants to learn Triton from scratch, onboard onto a Triton codebase, build a Triton exercise list, or preprocess Triton PDFs/docs into study material. Pair with the `triton-tutor` skill to quiz against the resulting vault.

Interactive quiz tutor for a Triton StudyVault built by `triton-tutor-setup`. Delivers 4-question rounds with concept-level proficiency tracking (🟥/🟨/🟩/🟦/⬜) across 6 Triton topics: Triton basics (`@triton.jit`, masks, `tl.load`/`tl.store`), tiling & autotuning (`triton.autotune`, `num_warps`/`num_stages`), matmul patterns (tiled GEMM, pid swizzling, persistent, split-K, FP8 `tl.dot`), attention & reductions (FlashAttention, online softmax, `tl.associative_scan`), compiler internals (TTIR/TTGIR/LLIR, NVIDIA/AMD backends, WGMMA + TMA lowering), and ecosystem/production (`torch.compile`/Inductor, AOT, `proton`). Use when the user wants to take a diagnostic Triton assessment, drill weak Triton concepts, study a specific Triton topic, review the learning dashboard, or says things like "quiz me on Triton", "drill FlashAttention", "/triton-tutor", "퀴즈".

Generate a university-style course pack (PDF lecture handouts, programming-assignment writeups, conceptual problem sets, topic cheatsheets, a capstone project, and a one-page syllabus) plus matching graded programming-exercise scaffolds (CUDA / Triton / PyTorch starters, a LeetGPU-style autograder harness, and gated solutions for hardest exercises) from a `StudyVault/` produced by `cuda-tutor-setup` or `triton-tutor-setup`. Visual style modeled on CMU 15-418, Stanford CS149, MIT 6.172, OLCF CUDA Training Series, and AlphaGPU LeetGPU challenges. Emits `.tex` + `Makefile`; compiles via `latexmk + pdflatex` if available, otherwise gracefully falls back to source-only. Use when the user wants to produce printable lecture handouts, programming assignments, problem sets, an autograder, or a complete course pack from an existing StudyVault, or says things like "make handouts for my StudyVault", "generate the course pack", "build programming exercises for CUDA", "/tutor-handouts".

Write, review, and revise academic Research Statements / Research Visions / Research Plans / 5-Year Plans for faculty, postdoc, fellowship, and promotion applications across disciplines and regions. Covers Mission/Strategy/Evidence/Story, US-school guidance from Cornell, CMU, MIT, Caltech, Penn, Yale, Harvard, UConn, Notre Dame, Delaware, Berkeley, length tiers, openings, past-work theming, fundable future lines, department fit, funding vectors, rubric scoring, red flags, and examples. Use when asked to write, draft, outline, edit, critique, score, red-team, or improve any research statement, research vision, research plan, future research statement, or research pitch.

Write academic grant proposals for major funders — NSF (CAREER, standard, CRII), NIH (R01, R21, K-series, F-series), ERC (Starting, Consolidator, Advanced), EU Horizon Europe, DARPA/DoD/DOE BAAs, and UKRI/EPSRC/Royal Society. Covers section drafting (Specific Aims, Project Description, Broader Impacts, Excellence/Impact/Implementation, Heilmeier Catechism), mock-reviewer critique against funder criteria, budget and budget justification, biosketches, Data Management Plans, and responses to reviewer critique on resubmission. Discipline focus: CS / Systems / ML, with examples aligned to compute-heavy research. Use when asked to draft, review, critique, revise, or budget any academic grant proposal, white paper, pre-proposal, or LOI.

Per-agent persistent memory for oh-my-open-code. Stores and retrieves knowledge across sessions using agent_id scoping — each agent (oracle, librarian, explore, sisyphus, deep, etc.) maintains its own memory namespace while allowing cross-agent search. Primary use case: persisting deep research findings, architecture decisions, library patterns, and codebase knowledge. Triggers on: "remember", "store this", "what do we know about", "recall", "search memory", "persist findings", or any need to retain knowledge across sessions.

Review systems conference papers for OSDI, NSDI, SIGCOMM, MOBICOM, SOSP, and FAST. Produces structured HotCRP-style reviews with paper summary, strengths, weaknesses, detailed comments, questions for authors, and overall merit/confidence scoring. Covers evaluation criteria (novelty, soundness, significance, evaluation quality, clarity, relevance, reproducibility), venue-specific expectations, common paper weaknesses, constructive feedback tone, one-shot revision reviews, and reviewer ethics. Use when asked to review, critique, or provide feedback on a systems conference paper submission.

Write systems conference papers for OSDI, NSDI, SIGCOMM, MOBICOM, and FAST. Covers paper structure (Introduction, Background/Motivation, Design, Implementation, Evaluation, Related Work), systems writing style, IEEE/ACM citations, architecture diagrams, performance evaluation, and artifact preparation. Includes FAST-specific guidance for short papers (6 pages) and deployed-systems papers.

Write conference paper rebuttals and author responses that effectively counter incorrect or unreasonable reviewer comments. Targets systems venues (OSDI, NSDI, SIGCOMM, MOBICOM, SOSP, ASPLOS, EuroSys, USENIX Security, CCS, PLDI) on HotCRP and ML venues (NeurIPS, ICML, ICLR, AAAI) on OpenReview. Core focus: identify and resolve reviewer false impressions with evidence-backed, firm-but-professional corrections. Covers rebuttal triage, false impression taxonomy (8 types with resolution playbooks), venue-specific constraints, response structure patterns, and ready-to-use phrase templates. Use when asked to write, draft, review, or improve a conference paper rebuttal, author response, or response to reviewer comments.

Transform raw meeting notes, transcripts, or recordings into actionable meeting summaries focused on decisions, action items with owners and deadlines, and accountability. Use when the user asks to summarize a meeting, extract action items from meeting notes, create meeting minutes, write a meeting recap, process a meeting transcript, or generate follow-up documentation from any meeting. Triggers on mentions of "meeting summary", "meeting notes", "meeting recap", "action items from meeting", "meeting minutes", "standup notes", "retro summary", or when pasted text appears to be a meeting transcript or notes.

Poll TickTick for coding tasks and dispatch them to remote machines via SSH + OpenCode Serve. Includes full TickTick CLI, intelligent agent routing (build/plan/deep), stuck/loop detection, plan verification gate, session knowledge management (DAG/forest), and LLM-powered session relevance matching with Haiku/GPT-5.1/GPT-4o fallback chain.

Bridge OpenClaw messaging channels (WhatsApp, Discord, Telegram, Slack, etc.) to Claude Code CLI. Use when the user wants to set up remote access to Claude Code through messaging apps, configure the OpenClaw-to-Claude bridge, manage cross-channel sessions, debug message routing, or format Claude responses for channel-specific limits.

AST-aware structural code search, rewriting, and linting via ast-grep. Use for structural search (find function calls, match patterns with meta-variables), code rewriting (rename functions, migrate APIs, convert patterns), custom lint rules (detect anti-patterns, enforce conventions), and any task where text-based grep is too imprecise for code structure.

Performs fast text search with one-shot patterns that minimize iterations by getting files, lines, and context in a single call. Use this skill to replace grep when searching for text patterns, finding specific code locations, or getting context around matches

Fetch BibTeX citations from DOI, arXiv ID, journal URL, or paper title. Covers identifier resolution, BibTeX normalization, batch citation collection, and integration with LaTeX bibliography workflows.

Create publication-quality matplotlib figures for two-column conference papers (ACM, IEEE, USENIX). Provides style presets, colorblind-safe Wong palette, and dual-output (PDF+PNG) workflow. Use when the user wants to create, style, or fix plots/figures/charts for academic paper submission, or asks for conference-quality matplotlib figures.

AST-accurate function dependency analysis using Tree-sitter. Use when tracing function call chains, finding callers/callees, analyzing blast radius of changes, detecting unused functions, or finding circular dependencies. Replaces grep-based code exploration with precise, structural matching. Supports Python, JavaScript, TypeScript, Go, Rust, C, C++, Java. Triggers on questions like "what calls X?", "what does X call?", "find unused functions", "show me the call graph", "what's the impact of changing X?", or any function dependency/relationship query.

Context database reference for OpenViking — an open-source context database for AI Agents that unifies memory, resources, and skills through a filesystem paradigm with hierarchical context delivery (L0/L1/L2) and semantic retrieval. Use this skill when the user asks about OpenViking, needs help writing code that uses the openviking Python package, wants to configure or troubleshoot OpenViking, or needs to understand its architecture, APIs, or concepts (Viking URI, context types, retrieval mechanisms, session management, etc.). Trigger on mentions of: "openviking", "viking://", "OpenViking", "context database", "AGFS", "VikingFS", "ovpack", or related API methods like find/search/add_resource/add_skill/session.