Field guide / sub-domain
43
sources in this field
Updated July 16, 2026
Current thesis
The shortest path to orientation.
Claude Code revenue runs at $2.5B annually with 4% of all GitHub commits. Anthropic is moving upstream into Platform: Claude Managed Agents (PaaS, $2.58 fulfillment cost per $1k service, Linear SDK), the Advisor Strategy (Opus advisor + Sonnet/Haiku executor as first-class pattern), Monitor tool (event-driven background scripts replacing polling), and official claude-code-setup plugin recommending hooks, skills, MCPs, subagents, automations. First-party documentation spans a 14-minute video guide and 33-page Skills guide. Skill distribution formalized: gstack packages dotfiles-for-AI, ByteRover unifies scattered knowledge into queryable index, claude-smart adds cross-project learning, self-improving skills grade against binary eval criteria. NVIDIA hosts ~80 free models via APIs (DeepSeek, Kimi, GLM). Enterprise adoption is real: Intercom built 13 plugins with 100+ skills, $1M+ customers 40x'd to 500+. GitHub plugin marketplace auto-installs across team instances. Evo (open-source) turns codebases into self-instrumenting research loops. Tolaria's MCP server lets Claude natively edit vaults (100K+ LOC, 85% test coverage). LibreChat's native MCP support extends ecosystem through self-hosted wrappers. Claude Ads audits 250+ checks, cuts waste 35%. Anthropic employees share internal methods for tracking Claude development publicly.
Evidence board
Claims worth carrying forward01Claude Teams supports a GitHub plugin marketplace pattern: host multiple plugins with sub-skills and agents in a central GitHub repo, configure auto-install on each team member's Claude instance, and push updates centrally—changes propagate to all users automatically.
02The GitHub plugin marketplace model solves the fleet synchronization problem for Claude at scale: one update to the central repository syncs to every team member's instance, making plugin lifecycle management a dev workflow rather than a per-user IT task.
03LLM Wikis + HTML Artifacts form a two-layer stack: wikis hold structured knowledge, artifacts render it interactively. Artifacts are bidirectional—they invoke agents and agents can update them—enabling automation workflows like inbox-zero, deep research, and experiment triggering without a full web app.
04HTML artifacts are not websites—they are lightweight interactive files that extend dynamically as needs grow. HTML and Markdown are complementary, not substitutes: Markdown for knowledge storage, HTML for interactive presentation and action surfaces on top of that knowledge.
05Hooking HTML artifacts bidirectionally to agents unlocks a personal work OS spanning: inbox-zero, topic monitoring, fast prototyping, deep research, experiment design/triggering, figure generation, research scheduling, and semantic search—all from a single artifact layer.
06Install the official Anthropic plugin via `/plugin install claude-code-setup@claude-plugins-official` to auto-scan your project and recommend hooks, skills, MCP servers, subagents, and automations—then set them up step-by-step. Most users running vanilla Claude Code miss this entirely.
07CLAUDE.md functions as persistent project memory across sessions: store architecture decisions, coding patterns, debugging notes, edge cases, and recurring mistakes. This eliminates re-explaining context every session and directly improves output quality.
08Run `/init` before touching any new codebase. Without it, Claude enters with near-zero project understanding; with it, Claude maps structure, dependencies, conventions, and patterns—output quality difference is immediate.
Adjacent fields
Latest evidence
Ruari Wood-Freeman
A GitHub-based plugin marketplace pattern allows centralized management of multiple Claude plugins with sub-skills and agents across an entire organization. Updates to the central repo automatically sync to all users' Claude instances, eliminating manual distribution overhead.
elvis
A workflow pattern combining LLM Wikis (structured knowledge stores) with interactive HTML Artifacts creates a bidirectional agent-UI layer. Artifacts are built on top of wikis, can invoke agents, and agents can update artifacts—enabling inbox-zero automation, research scheduling, topic discovery, and live figure generation from a single HTML file.
Suryansh Tiwari
Claude Code's real power comes from building the ecosystem around it—persistent memory (CLAUDE.md), project initialization (/init), parallel Git worktrees, MCP servers, plugins, subagents, and CI/CD integration—rather than prompting better. The claude-code-setup plugin automates recommending and installing these components.
Claude
Claude Code has launched 'agent view,' a consolidated list of all active sessions, currently available as a research preview. The feature addresses session management as Claude Code's multi-agent usage scales.
Yi Lu
claude-smart is an open-source plugin for Claude Code, Codex, and OpenCode that turns corrections and successful patterns into structured, reusable skills — stored locally at ~/.reflexio/ — rather than passive memory transcripts. It claims ~3× better rule retention than claude-mem, 70%+ reduction in planning iterations on repeated tasks, and runs entirely on-device with ONNX embeddings (all-MiniLM-L6-v2) plus hybrid vector+BM25 search.
Dan McAteer
'Understand Anything' is an open-source plugin for Claude Code, Codex, and OpenCode that analyzes a directory and builds an interactive knowledge base explaining structure rather than showing it. Notably, it generalizes beyond code — running /understand on an 888-episode Obsidian podcast vault (144K lines of markdown) produced a searchable knowledge graph.
Sean Geng
Sean Geng packaged @goodalexander's iterative plan-improvement technique as an installable Claude Code skill. The skill treats planning as a search problem: generate a plan, score it 0-100 against a rubric, critique specific weaknesses, rewrite, keep the best, and stop when scores plateau over a sliding window. One curl command installs it into ~/.claude/skills/plan-optimizer/.
Matthew Berman
Loop Library (signals.forwardfuture.com/loop-library) is a searchable collection of 85+ reusable AI agent loop prompts organized by domain (Engineering, Evaluation, Operations, Content, Design). Each loop includes explicit stopping conditions, verification steps, and proof requirements—addressing the core failure mode of agents that report partial work as complete.
Samuel Stroschein
Flashtype is an open-source markdown editor that integrates Claude and Codex to edit markdown files, track every change, and display diffs. It targets developers who want AI-assisted document editing with full version visibility.
Nabeel Hyatt
Nabeel Hyatt describes prompting Fable to read relevant academic literature on an idea before building, then think adversarially — a two-step research-then-critique pattern that front-loads domain knowledge and stress-testing before implementation begins.