meta-plugin

The Builder for an agent-team run — the only role that edits. Use to produce the work product (code, docs, research, or design) against an approved plan and explicit acceptance criteria. Owns only the files the spawn prompt assigns; self-verifies and shows the evidence.

The adversarial Critic (Devil's-Advocate) for an agent-team run. Use to review a diff/output against acceptance criteria with fresh context — given only the work and the criteria, never the Builder's reasoning. Read-only; flags correctness/requirement defects with severity and does not over-report.

The gate-keeping Evaluator for an agent-team run. Use to score a work product against each explicit acceptance criterion and emit PASS/FAIL backed by real evidence (command/test output), kept distinct from the Critic. Read-only; reports failures, never fixes them.

The Integrator for an agent-team parallel slice build. Use to assemble interdependent slices produced by parallel Builders against the pinned contract, resolve integration seams, and run the integration checks with real output as evidence. Edits only the seams — not slice internals.

The DEFAULT review role for an agent-team run. A single fresh-context agent that both finds correctness/requirement defects (each with a severity) and renders the PASS/FAIL gate verdict backed by real evidence — given only the diff/output and the acceptance criteria. A non-blocking flaw still passes. Read-only. (Split into a separate Critic + Evaluator only for high-stakes work.)

Use when the user invokes /agent-team <task>, OR when a task is substantial enough to warrant a persistent, role-based agent team the user drives as lead — multi-file features, large refactors or migrations, deep multi-source research, or design work where a gated refine → review loop beats a one-shot answer. Designs the team's topology to fit the task (research fan-out · gated build/refine · interdependent parallel slices · staged pipeline) and scaffolds single-responsibility teammates (Scout, Builder, Reviewer — split into Critic + Evaluator for high-stakes work — plus Synthesizer/Integrator as needed) backed by real subagent definitions that enforce per-role tool whitelists and read-only/edit boundaries, each spawned with a full 4-part contract; the lead orchestrates and synthesizes but never authors the work product, stops at two human gates (plan, then ship), caps the loop, and never ships autonomously. Use even if the user doesn't name the command. Do NOT use for a single quick delegation, a one-shot parallel fan-out, a pure question, or a trivial single-file edit a lone agent handles faster — a team's ~7× token cost only pays off when the work is high-value, parallelizable, or larger than one context window.

Manually-invoked skill for generating PRD.md (what + why, tech-agnostic), SPEC.md (architecture + alternatives + rollout + cross-cutting concerns), and CLAUDE.md (commands, stack, conventions, gotchas for Claude Code sessions). Interviews to fill gaps, dispatches an inventory subagent on brownfield repos, applies a hard pre-flight quality gate per doc, emits one, two, or all three on request.

Use when the user shares an architecture, system design, tech blog, code, or implementation detail and wants the foundational system-design concepts behind it explained or broken down from first principles for their own learning — e.g. "explain the concepts behind this", "break this design down", "help me understand why this works", "what's actually going on here".

Use when the user invokes /vault-save (optionally with a subfolder name like 'claude-notes', 'claude-research-output', or 'meta-workflows'), OR when the user asks to save, export, copy, or stash the latest response into their Obsidian vault. Captures the most recent prior assistant response as a markdown file with YAML frontmatter (date, source=current git branch/worktree, tags) under the Obsidian vault at /Users/hoangngo/Documents/personal-vault/, fuzzy-matches the subfolder argument against existing folders, and prompts for confirmation before writing.