Collection of Artificial Intelligence / LLM resources I actually use day to day. Biased towards Claude and the Ruby ecosystem, because that’s where I spend most of my time.
Claude Code
Claude Code is Anthropic’s agentic coding CLI. It lives in the terminal, reads and edits files, runs commands, and keeps track of project context across sessions. Paired with CLAUDE.md files in the repo root, it becomes a pretty capable pair programmer.
- CLI reference
- Interesting uses, e.g. let Claude generate a presentation of your codebase with Marp
- Roadmap management with Claude Code
- Building Clara: a Claude Code setup for running a consultancy — my own workflow: PARA-style markdown repo, custom persona, MCP servers and a few cron jobs.
Subagents & hooks
Two building blocks I lean on heavily in Clara:
- Subagents — specialised sessions, each with its own context window and tool access. Claude hands a side task (research, a broad search) to one and gets back just the summary, so the main conversation stays clean. You can also pin them to a cheaper model like Haiku.
- Hooks — shell commands that fire on lifecycle events (before/after a tool runs, on session start, …). The place to enforce conventions, auto-format, or wire in your own automation without nagging the model to remember.
Plugins & extensions
Plugins bundle skills, subagents, hooks and MCP servers into one installable unit. The /plugin marketplace is the central place to browse and install them — Anthropic ships a curated official marketplace, and you can add community ones (any GitHub repo or Git URL).
- Codex plugin for Claude Code — makes OpenAI’s Codex available from inside Claude Code. Handy for a second opinion (
/codex:review,/codex:adversarial-review) or for delegating background tasks (/codex:rescue). - Claude for Legal — Anthropic’s plugin suite for legal workflows (commercial, corporate, employment, privacy, IP, litigation, regulatory, AI governance). 100+ named agents that draft, triage and review documents, with MCP connectors for CourtListener, Westlaw, Ironclad, iManage and others. Every output is framed as a draft for attorney review. A good template for what a domain-specific Claude Code plugin pack can look like.
Spec-driven workflows
Both of these treat specifications as the source of truth: you write the spec first, then agents implement against it. Different flavours for different project sizes.
- GitHub Spec Kit — GitHub’s own SDD toolkit and the de-facto standard (90k+ stars). The flow is Spec → Plan → Tasks → Implement via
/specify,/planand/tasksslash commands. Agent-agnostic: it drives Claude Code, Copilot, Gemini and ~30 others, so you can switch the underlying agent without rewriting the workflow. - CCPM – Claude Code Project Management — uses GitHub Issues and Git worktrees to let up to 12 Claude instances work in parallel. Workflow is Brainstorm → PRD → Epic → Issues → Tracking, all via
/pm:*slash commands. - Spec Kitty — CLI plus live Kanban dashboard, 6-phase lifecycle (spec → plan → tasks → implement → review → merge). Supports 12 agents including Claude Code, Cursor, Windsurf, Gemini and GitHub Copilot.
Agent Skills
Skills are self-contained capability packages — a folder of instructions, scripts and resources that Claude can load on demand.
Model Context Protocol (MCP)
MCP is the open protocol that lets LLMs talk to external tools and data sources in a standardised way. The magic is that any MCP-compatible client (Claude Desktop, Claude Code, Cursor, …) can use any MCP server.
MCP servers I rely on:
- Linear MCP — read and update Linear issues from within Claude.
- Sunsama MCP — full CRUD on Sunsama tasks, subtasks and streams. Great for keeping the daily plan in sync without leaving the terminal.
- Member Berries MCP — built on apple-mcp, adds memory-style tools for calendar, notes and reminders. Turns Claude into “a helpful friend who remembers your day”.
Ruby/Rails tools and frameworks
- RubyLLM — one clean Ruby API across OpenAI, Anthropic, Gemini, Ollama and friends: chat, vision, embeddings, tool calls and streaming, with a model registry covering hundreds of models. First-class Rails integration (
acts_as_chat, install generators), so wiring an LLM into an app is genuinely a handful of lines. My default starting point these days. - Raif (Ruby AI Framework) — a Rails engine that helps you add AI-powered features to your Rails apps.
- Claude on Rails — Obie Fernandez’s framework that orchestrates specialised agents (Architect, Models, Controllers, Views, Services, Tests, DevOps) on top of
claude-swarm. Rails-aware, test-driven, integrates with the Rails MCP Server for up-to-date documentation. - classifier-reborn — a general classifier module for Bayesian and other types of classification. Pre-LLM, but still useful for lightweight text categorisation without calling an API.