Agent Skills

---

name: modern-python-toolchain

description: Modern Python project setup with uv, ruff, and pyright. Use when initializing a new Python project, configuring the Python environment, setting up linting/formatting, or when a project needs uv (the fast Python package manager). Trigger on: 'set up Python', 'new Python project', 'configure uv', 'install uv', 'ruff', 'pyright', 'Python linting', 'Python formatting', or when a task requires Python and no pyproject.toml exists yet.

---

name: pdf-official

description: Use this skill whenever a PDF file is being produced, opened, transformed, filled, or read. That includes: extracting text or tables from an existing PDF; combining, carving, rotating, cropping, or watermarking pages; composing a fresh PDF (report, invoice, certificate); filling AcroForm fields or overlaying text onto a non-fillable scanned form; encrypting or unlocking a PDF; running OCR over a scanned document; rendering pages to PNG/JPEG for visual analysis. Trigger on mentions of 'PDF', a filename ending in .pdf, requests like 'turn this into a PDF report', or references to AcroForm / form fields.

---

name: playwright

description: Use when the task requires automating a real browser from the terminal (navigation, form filling, snapshots, screenshots, data extraction, UI-flow debugging) via `playwright-cli` or the bundled wrapper script.

---

name: pptx-official

description: Use this skill whenever a Microsoft PowerPoint (.pptx) file is being produced, opened, transformed, or read. That includes: authoring slide decks, pitch decks, executive readouts, training material, or any presentation deliverable; extracting text or structure from an existing .pptx; filling a .pptx template with values; converting a deck to PDF or images; splitting or merging decks; inspecting slides, layouts, masters, tables, images, charts, speaker notes, or comments. Trigger on words like 'deck', 'slides', 'presentation', 'pitch deck', 'keynote' (when a .pptx is expected as output), or any filename ending in .pptx. Do NOT trigger when the primary deliverable is a Word document, spreadsheet, PDF report, HTML site, or Google Slides API call, even if presentation-shaped content appears along the way.

---

name: product-design

description: Use this skill when Product Design is explicitly invoked or the main task is product design exploration, UX research, flow auditing or critique, visual ideation, cloning a live product surface, implementing a selected visual target, design QA, saved design context, or sharing a prototype.

---

name: research-paper-writing

description: Write, rewrite, and polish academic papers (ML/CV/NLP style). Use when the user drafts or revises Abstract, Introduction, Related Work, Method, Experiments, or Conclusion; asks "does this flow / 这段通顺吗 / polish this paragraph"; turns bullet points or a Chinese draft into publication-quality English; runs a pre-submission self-review or reviewer-style critique; fixes paper figures/tables/LaTeX formatting; or compiles/converts the paper to PDF (LaTeX build, 编译PDF, 转成PDF). Trigger on mentions of paper, draft, camera-ready, rebuttal-facing revision, CVPR/ICCV/NeurIPS/ICLR/ACL-style venues, or .tex files being edited for a paper.

---

name: sales

description: Use this skill whenever Sales is explicitly invoked or the task involves customer meeting preparation, call follow-up, account prioritization, account signals, deal strategy, business cases, competitive briefs, forecasts, customer evidence, rep coaching, company research, CRM context, or company and contact enrichment.

---

name: skill-creator

description: Interactive guide for creating, reviewing, and improving agent skills (SKILL.md folders). Use when the user wants to build a new skill ('create a skill', 'make a skill for X', 'write a SKILL.md', 'turn this workflow into a skill'), review or improve an existing skill, fix a skill that never triggers or triggers too often, or validate a skill folder before sharing it. Do NOT use for general prompt writing, MCP server development, or editing arbitrary markdown files.

---

name: super-research

---

name: xlsx-official

description: Spreadsheet toolkit. Reach for it whenever the artifact on either side of the conversation is a workbook file — .xlsx, .xlsm, .xltx, .csv, .tsv — and the user wants that artifact produced, changed, cleaned, or read. Typical triggers: 'build me a model', 'update this sheet', 'add a column', 'compute the totals as formulas', 'sanity-check this xlsx', 'export sheet 2 to CSV', 'render the workbook as PDF', 'the spreadsheet in ~/Downloads is a mess, fix it'. Applies equally to financial models, ops reports, data cleanups, and template fills. Skip it when the workbook is only source material and the real output is a Word doc, an HTML page, a Python script that runs standalone, a Google Sheets integration, or an ingestion pipeline into a database — in those cases the spreadsheet is a means, not the deliverable.

---

name: autoresearch

description: Orchestrates end-to-end autonomous AI research projects using a two-loop architecture. The inner loop runs rapid experiment iterations with clear optimization targets. The outer loop synthesizes results, identifies patterns, and steers research direction. Routes to domain-specific skills for execution, supports continuous agent operation via Claude Code /loop and OpenClaw heartbeat, and produces research presentations and papers. Use when starting a research project, running autonomous experiments, or managing a multi-hypothesis research effort.

---

name: axolotl

description: Expert guidance for fine-tuning LLMs with Axolotl - YAML configs, 100+ models, LoRA/QLoRA, DPO/KTO/ORPO/GRPO, multimodal support

---

name: constitutional-ai

description: Anthropic's method for training harmless AI through self-improvement. Two-phase approach - supervised learning with self-critique/revision, then RLAIF (RL from AI Feedback). Use for safety alignment, reducing harmful outputs without human labels. Powers Claude's safety system.

---

name: distributed-llm-pretraining-torchtitan

description: Provides PyTorch-native distributed LLM pretraining using torchtitan with 4D parallelism (FSDP2, TP, PP, CP). Use when pretraining Llama 3.1, DeepSeek V3, or custom models at scale from 8 to 512+ GPUs with Float8, torch.compile, and distributed checkpointing.

---

name: fine-tuning-with-trl

description: Fine-tune LLMs using reinforcement learning with TRL - SFT for instruction tuning, DPO for preference alignment, PPO/GRPO for reward optimization, and reward model training. Use when need RLHF, align model with preferences, or train from human feedback. Works with HuggingFace Transformers.

---

name: grpo-rl-training

description: Expert guidance for GRPO/RL fine-tuning with TRL for reasoning and task-specific model training

---

name: huggingface-tokenizers

description: Fast tokenizers optimized for research and production. Rust-based implementation tokenizes 1GB in <20 seconds. Supports BPE, WordPiece, and Unigram algorithms. Train custom vocabularies, track alignments, handle padding/truncation. Integrates seamlessly with transformers. Use when you need high-performance tokenization or custom tokenizer training.

---

name: implementing-llms-litgpt

description: Implements and trains LLMs using Lightning AI's LitGPT with 20+ pretrained architectures (Llama, Gemma, Phi, Qwen, Mistral). Use when need clean model implementations, educational understanding of architectures, or production fine-tuning with LoRA/QLoRA. Single-file implementations, no abstraction layers.

---

name: llama-factory

description: Expert guidance for fine-tuning LLMs with LLaMA-Factory - WebUI no-code, 100+ models, 2/3/4/5/6/8-bit QLoRA, multimodal support

---

name: llamaguard

description: Meta's 7-8B specialized moderation model for LLM input/output filtering. 6 safety categories - violence/hate, sexual content, weapons, substances, self-harm, criminal planning. 94-95% accuracy. Deploy with vLLM, HuggingFace, Sagemaker. Integrates with NeMo Guardrails.

---

name: mamba-architecture

description: State-space model with O(n) complexity vs Transformers' O(n²). 5× faster inference, million-token sequences, no KV cache. Selective SSM with hardware-aware design. Mamba-1 (d_state=16) and Mamba-2 (d_state=128, multi-head). Models 130M-2.8B on HuggingFace.

---

name: miles-rl-training

description: Provides guidance for enterprise-grade RL training using miles, a production-ready fork of slime. Use when training large MoE models with FP8/INT4, needing train-inference alignment, or requiring speculative RL for maximum throughput.

---

name: nanogpt

description: Educational GPT implementation in ~300 lines. Reproduces GPT-2 (124M) on OpenWebText. Clean, hackable code for learning transformers. By Andrej Karpathy. Perfect for understanding GPT architecture from scratch. Train on Shakespeare (CPU) or OpenWebText (multi-GPU).

---

name: nemo-curator

description: GPU-accelerated data curation for LLM training. Supports text/image/video/audio. Features fuzzy deduplication (16× faster), quality filtering (30+ heuristics), semantic deduplication, PII redaction, NSFW detection. Scales across GPUs with RAPIDS. Use for preparing high-quality training datasets, cleaning web data, or deduplicating large corpora.