Claude Skills by anubhavg-icpl

Production self-host Ollama in Docker/Compose with GPU passthrough, model preload, reverse proxy auth, and multi-GPU. Use when deploying, running, or configuring local LLM inference with ollama docker deploy.

Run, customize, and serve local LLMs with Ollama, Modelfiles, and GGUF quantization. Use when building AI applications with ollama.

Publish models to ollama.com/library — namespace setup, ollama push, signing keys, quant tags, parameter-size tags, model card README authoring. Use when creating, converting, or publishing model files with ollama library publisher.

Author production Modelfiles with FROM, PARAMETER, TEMPLATE, SYSTEM, ADAPTER, MESSAGE, and LICENSE directives for Llama 3, Qwen, Phi, and Gemma. Use when creating, converting, or publishing model files with ollama modelfile.

Author Ollama Modelfiles for vision models — llava, llama3.2-vision, MiniCPM-V — with mmproj projector handling and image-token templates. Use when creating, converting, or publishing model files with ollama multimodal modelfile.

Size a Thinker-Talker streaming voice pipeline (Qwen-Omni / Moshi / Mini-Omni) for a target TTFAB and feature set. Use when you need help with omni streaming budget.

Allocate LLaVA-OneVision-style unified visual-token budgets across single-image, multi-image, and video scenarios for a target product mix. Use when you need help with onevision budget planner.

>. Use when you need help with open design landing deck.

>. Use when you need help with open design landing.

Pick an open LLM family, quantization, and inference stack for a given deployment target. Use when you need help with open model picker.

Build production agents with handoffs, guardrails, and tracing using the OpenAI Agents SDK. Use when building AI applications with openai agents sdk.

Expert in the OpenAI design system - Calm, near-monochrome system anchored in deep teal-black with generous white space and editorial typography. Use when building UI components, applying design tokens, or implementing visual styles for openai.

openai/evals framework — registry layout, model-graded patterns, custom YAML evals. Use when evaluating, monitoring, or observing LLM performance with openai evals.

Expert in the OpenCode design system - AI coding platform. Developer-centric dark theme. Use when building UI components, applying design tokens, or implementing visual styles for opencode ai.

Deep expertise in OpenSearch k-NN — Lucene/Faiss/NMSLIB engines, neural sparse, hybrid query DSL, and ML Commons inference. Use when implementing vector search, embeddings storage, or similarity queries with opensearch vector.

Expert in OpenTelemetry for distributed tracing, metrics, and logging. Use when configuring, deploying, or managing opentelemetry infrastructure.

OpenTelemetry GenAI semantic conventions — standardized LLM spans for any vendor. Use when evaluating, monitoring, or observing LLM performance with opentelemetry llm.

|. Use when you need help with orbit general.

|. Use when you need help with orbit github.

|. Use when you need help with orbit gmail.

|. Use when you need help with orbit linear.

|. Use when you need help with orbit notion.

Pick an orchestration topology (supervisor, swarm, hierarchical, debate, or none) for a given problem and implement it minimally. Use when you need help with orchestration picker.

Odds-Ratio Preference Optimization — single-stage SFT + preference alignment without a reference model. Use when fine-tuning, training, or adapting language models with orpo techniques.

Produce an instrumentation plan for an agent codebase to emit OTel GenAI spans end-to-end. Use when you need help with otel genai instrumentation.

Instrument an agent with OpenTelemetry GenAI semantic conventions — invoke_agent, chat, tool_call spans with correct attributes and opt-in content capture. Use when you need help with otel genai.

Guarantee structured LLM outputs with regex, JSON schema, and grammar-constrained generation. Use when building AI applications with outlines.

Decision framework for choosing the right LLM evaluation strategy based on task type, budget, and requirements. Use when you need help with skill evaluation.

Pick tokenizer algorithm, vocab size, library for a given corpus and deployment target. Use when you need help with tokenizer picker.

Expert in the Pacman design system - Retro arcade-inspired design with pixel fonts, dotted borders, playful high-contrast colors, and 8-bit game aesthetics. Use when building UI components, applying design tokens, or implementing visual styles for pacman.

Interactive pair programming mentor that teaches while coding together. Use when you need help with pair programming mentor.

Expert in the Paper design system - Paper-textured, print-inspired design with minimal colors, clean serif/sans typography, and tactile surface qualities. Use when building UI components, applying design tokens, or implementing visual styles for paper.

Audit a tool registry for safe parallelization. Mark each tool parallel_safe, note ordering dependencies, and flag downstream rate-limit risk. Use when you need help with parallel call safety check.

Route a reasoning workload between voting, tree-of-thought, multi-agent, Hogwild!, and speculative decoding strategies. Use when you need help with parallel inference router.

Small chunks for retrieval, large parents for context — sentence-window, multi-vector. Use when building or optimizing retrieval-augmented generation pipelines with parent child retriever.

Read a ViT config and produce a patch-token, parameter, and VRAM analysis for downstream VLM planning. Use when you need help with patch geometry reader.

Expert in PCI-DSS compliance for payment card security - cardholder data protection, network security, and audit controls. Use when performing security analysis, auditing, or hardening with pci dss compliance.

HuggingFace PEFT library survey — LoRA, IA3, prompt tuning, prefix tuning, AdaLoRA, OFT/BOFT, VeRA. Use when fine-tuning, training, or adapting language models with peft techniques.

Expert in debugging performance issues, bottlenecks, and optimization. Use when diagnosing, troubleshooting, or fixing bugs with performance debugging.

Web performance optimization including bundle analysis, lazy loading, caching strategies, and Core Web Vitals. Use when you need help with performance optimization.

performance-testing. Use when writing, running, or improving tests with performance testing.

Match a Claude Code task to the correct permission mode, budget caps, and required isolation before starting a run. Use when you need help with permission mode picker.

Expert in the Perspective design system - Spatial depth design with isometric views, vanishing points, and layered elements that guide attention through 3D-like realism. Use when building UI components, applying design tokens, or implementing visual styles for perspective.

Deep expertise in pgvector 0.8+ for PostgreSQL — HNSW/IVFFlat tuning, halfvec/sparsevec, hybrid search with tsvector + RRF, and pgvectorscale (DiskANN). Use when implementing vector search, embeddings storage, or similarity queries with pgvector.

Expert in Phoenix - Elixir's productive web framework for reliable, fast applications. Use when building applications with the phoenix framework.

Production-ready Phoenix/Elixir project structure with contexts, LiveView, and OTP patterns. Use when scaffolding, structuring, or architecting phoenix projects.

Autonomous agent that audits PHP/Laravel codebases for security vulnerabilities based on OWASP and RFC standards. Use when performing security analysis, auditing, or hardening with php laravel security audit agent.

php-laravel. Use when writing, reviewing, or refactoring php laravel code.

Web search and content extraction via Brave Search API. Use for searching documentation, facts, or any web content. Lightweight, no browser required.

Interactive browser automation via Chrome DevTools Protocol. Use when you need to interact with web pages, test frontends, or when user interaction with a visible browser is required.