agents-architect

Author a single sub-agent from a short brief. Use to add a specialist to an existing plugin.

Author a single slash command from a brief. Use to extend an existing plugin with a new /command.

Checkpoint the current agentic session to disk and emit a resume prompt so you can /compact safely. Use when context has degraded (tier DEGRADING or POOR) mid-workflow.

Install a generated plugin locally via the marketplace mechanism. Self-installs agents-architect or installs any plugin produced by /agents-architect:new.

Iterate on an existing agents-architect-generated plugin — re-run eval, fix flagged skills/agents/commands, bump version. Use after /agents-architect:new when the eval report suggests improvements, or after manual edits.

Writes sub-agent definition files (agents/*.md) for a new plugin given a DOMAIN.md and an agent plan. Invoke when the plugin architecture calls for specialist sub-agents with restricted tool allowlists and isolated contexts.

Writes slash-command files (commands/<ns>/<cmd>.md) for a new plugin given a COMMAND_PLAN. Invoke when the plugin needs user-facing /commands that orchestrate skills and sub-agents into repeatable workflows.

Designs the context management layer for a new plugin — CLAUDE.md, memory tiers, scratchpad schemas, read-depth rules, compaction strategy. Invoke after skills/agents/commands are planned so the context layer reflects the full workflow graph.

Researches a domain end-to-end and produces DOMAIN.md for a new plugin build. Invoke when the user runs /agents-architect:new <domain> or /agents-architect:research <domain>, when the orchestrator needs a cited survey of tools/MCPs/practices/anti-patterns before scaffolding skills.

Runs the eval harness against a newly generated plugin — skill trigger accuracy, agent delegation correctness, command workflow integrity, regression suite. Invoke after /agents-architect:new packaging and whenever a plugin is iterated, to score and report.

Write or refactor Claude Code sub-agent definitions (files in agents/*.md). Use when the user asks to "create a subagent", "add an agent for X", "build a specialist agent", "make a research agent", "write an agent file", "agent YAML", "delegate to a subagent", "design a multi-agent workflow", or when a task needs a fresh context window, restricted tools, or isolation. Covers frontmatter (tools/model/effort/maxTurns/isolation), delegation patterns, read-only constraints, and orchestrator↔specialist contracts.

Write or refactor Claude Code slash commands (files in commands/<ns>/<cmd>.md). Use when the user asks to "create a slash command", "add /<name>", "write a command file", "make a /command", "turn this workflow into a command", "argument hints", "allowed-tools", or when packaging a repeatable workflow as a one-keystroke invocation. Covers frontmatter, $ARGUMENTS parsing, @-imports for context files, and orchestration patterns.

Design and optimize context windows for agentic workflows. Use when the user asks to "manage context", "optimize context", "compact context", "context budget", "CLAUDE.md", "memory hierarchy", "context degradation", "agent is forgetting", "agent hallucinating after many turns", "long conversation", "hit token limit", "prune context", "memory tiers", or when a multi-turn workflow shows quality decay. Covers CLAUDE.md authoring, tiered memory, scratchpads, /compact strategies, read-depth rules, and subagent delegation for context relief.

Research a domain end-to-end before scaffolding an agentic system for it. Use when the user asks to "research [domain]", "learn about [field]", "what are the best practices for X", "survey the tools for Y", "what MCPs exist for Z", "before we build the agent, understand [topic]", or when kicking off a new /agents-architect:new for an unfamiliar domain. Covers methodology: taxonomy → tools → MCPs → idioms → anti-patterns → eval cases. Produces DOMAIN.md.

Select, scaffold, or integrate Model Context Protocol (MCP) servers with a Claude Code plugin. Use when the user asks to "add an MCP", "which MCP for X", "connect to [service]", "find a server for [API/tool]", "wrap [tool] as an MCP", ".mcp.json", "MCP server config", or when a domain needs external tools Claude doesn't have (Blender, Figma, ClickUp, arxiv, a CAD program, etc.). Covers MCP discovery via registry, config schema, env/userConfig wiring, and fallback to CLI when no MCP exists.

2 hooks across 2 events