Claude

Claude Code is now a $2.5B run-rate product powering 4% of all GitHub commits, and Anthropic has moved decisively up the stack into Platform: Claude Managed Agents (PaaS for agents at $2.58 fulfillment cost on $1k of service, Linear SDK integration), the Advisor Strategy (Opus advisor + Sonnet/Haiku executor as a first-class platform pattern), the Monitor tool (event-driven background scripts replacing polling loops), /ultraplan (cloud-based planning with browser review and back-teleport to local CLI), Claude Design (Opus 4.7 vision rendering HyperFrames videos in 2 prompts), and official setup/plugin flows all shipped in this period. The workflow has matured into a general work operating system: CEOs use it as an AI Chief of Staff, Marcus Moretti runs Spiral at Every as a one-person PM/code/support/marketing team with strategy.md + a /ce:product-pulse cron replacing 60% of the old PM week, and outbound sales teams run 11-API pipelines through Skills files. Role specialization is now packaged (gstack installs a 6-specialist team — CEO/Eng Manager/Designer/Release Manager/Doc Engineer/QA Lead — in 30 seconds) and the loop is going multi-model (Opus 4.7 plans, GPT-5.5 reviews, Playwright validates, @conductor_build switches models for ~$400/mo). Plan files, /handover, structured /goal prompts, the premortem flip ("it's 6 months from now and this is already dead"), and the /ss screenshot skill all live alongside specialized harnesses (designer, marketer, sales, motion designer, bookkeeper), while a new agent-view research preview consolidates all coding sessions cross-project and a one-command skill promotes a personal PM OS into a team OS without leaking personal context.

The architecture is plain text markdown as a local knowledge base + Claude Code as the engine. Configuration is converging on three files for articulate agents: SOUL.md (constitution — voice/values, brutally specific or output reverts to ChatGPT), USER.md (~4000-word user model), AGENTS.md (operational playbook). Cost architecture matters: 80% of agent tasks are janitorial, making hierarchical model routing (10x cost reduction) essential. The scratchpad/napkin pattern provides a distinct memory form that compounds across sessions. Obsidian + Claude Code is the recognized community stack, and the vault pattern now extends to git-as-version-history (Tolaria) and scales to teams — four independent implementations (DoorDash, Pendo, Google, solo) converged on the same three-layer team-knowledge architecture, confirming that compounding data, not technology, is the moat.

Settings and configuration span six distinct extension mechanisms (Plugins, Skills, MCPs, Commands, Subagents, Hooks) across three layers: environment (--dangerously-skip-permissions, game sound hooks, performance-recovery env flags like CLAUDE_CODE_DISABLE_1M_CONTEXT), context architecture (CLAUDE.md under 200 lines, tiered manifests, progressive disclosure), and skill design (state machines, self-improving eval loops). System engineering beats prompt engineering.

Design now extends from the 3-layer harness (Skills + Canvas + Inspiration) to first-party Claude Design + HyperFrames (motion design via 2 prompts), Codex image-generation replacing UI prototyping, Refero's 2,000 DESIGN.md training files, and Lamina Labs' whiteboard animation SDK ("to draw" as agent primitive). The single-file markdown design spec is now a recognized pattern — Google's open-source Design.md format, best populated by having Claude/ChatGPT extract an existing brand's design language, then referenced by modular per-touchpoint skills for a unified brand.

The ecosystem is backed by Anthropic's own investments (11 open-source plugins, free course + 14-min agent guide + 33-page Skills guide, Managed Agents platform, official setup plugin) alongside community tooling: gstack skill packs, self-improving skills (32/50 → 47/50 overnight), claude-smart's local cross-project learning, Evo's autoresearch-as-plugin loop, team-scale plugin distribution (GitHub plugin marketplace auto-syncing across Claude Teams instances), CodexBar for token tracking, Decode and the screenshot skill for visual eyes, FieldTheory for X-bookmarks-as-context, Excalidraw skills for architecture diagrams, ByteRover for unified knowledge search, MCP-as-default knowledge tools (Tolaria), NVIDIA's free 80-model API, LibreChat-style self-hosted model wrappers, and 1,200+ hour research workflows. Voice cloning at 94% accuracy enables delegated communication. Codex as overflow when credits run low.