The five Claude Code tools that change frontend design work: impeccable skill, Figma MCP, claude-in-chrome MCP for visual QA, banana skill for AI image generation, and shadcn/ui MCP.
The five Claude Code tools that change how SEO and content teams work: seo-content skill, seo-technical skill, Firecrawl MCP, Exa MCP, and the filesystem MCP for direct CMS access.
The five Model Context Protocol servers worth installing today: filesystem, GitHub, Postgres, claude-in-chrome (browser), and Sentry. With install commands, the tools they expose, and the security model.
Five Claude Code skills that turn it from a chat tool into a development environment: impeccable for UI design, playwright-cli for browser automation, gsd for project planning, seo-content, and banana for image generation.
Write LLM evals that catch real regressions: pick the right metrics (exact match, LLM-as-judge, embedding similarity), build a golden dataset, run on every PR, and watch the trend over time.
Build an LLM agent with tool use: the agentic loop, the tool-call format on Anthropic, OpenAI, and Gemini, runnable code in JavaScript and Python, plus the common failure modes.
Add semantic search to an existing MySQL app with MySQL 9's VECTOR type, an embedding model (Voyage, OpenAI), and a cosine-similarity index. No separate vector database needed.
Pick the right Claude tier for the job: Haiku for high-volume cheap, Sonnet for the default workhorse, Opus for hard reasoning. With cost math, latency, and a decision matrix.
Cut LLM API costs by 90% with prompt caching on Anthropic, OpenAI, and Gemini. The math, what to cache, the TTL trade-offs, and code in JavaScript and Python.
Build RAG with embeddings, chunking, and vector search. Covers OpenAI / Voyage / Cohere embeddings, chunk sizes, hybrid search, reranking, and when long context replaces RAG entirely.
Run a local LLM with Ollama: install, pull a model, the hardware floor, picking between Llama, Mistral, and Qwen, and when local is faster than cloud (and when it isn't).
Get reliable JSON out of an LLM with native structured-output modes (Anthropic tool use, OpenAI Structured Outputs, Gemini schema), plus Zod / Pydantic validation as a fallback.
Six techniques that actually reduce LLM hallucination: grounding with retrieved context, citation requirements, tool use for facts, structured outputs, explicit don't-know permission, and LLM-as-judge verification.
Write an effective system prompt for an LLM with five parts: role, capabilities, constraints, output format, refusal policy. With before/after examples and the structure that maximises cache hits.