Produces and maintains a production-grade design doc system BEFORE code, in Alon's house format: SOURCE_OF_TRUTH.md, ARCHITECTURE_ROADMAP.md, TODO_WORKFLOW.md, and a CLAUDE.md rules file (auto-loaded by Claude Code), plus a modular docs/architecture/ set for larger projects. Model first — data + invariants → enforcement at the core → failure paths → API contracts → NFRs → phased workstream. All docs stay synced top-down at all times (docs before code). Design only, no implementation. Use when the user says "architect", "design doc", "spec this out", "before we build", "production-grade plan", "architecture for", "roadmap", "source of truth", or is starting a new app, feature, or workstream.
Convene a panel of opinionated advisors (parallel Claude subagents, each with a distinct mandate and a forbidden move so they genuinely diverge), gather independent positions, optionally cross-critique, then deliver a Chairman synthesis that COMMITS to one recommendation with explicit tradeoffs. For high-stakes judgment, design, and tradeoff calls — not fact-finding (use deep-research for facts). Use when the user says "ask the council", "convene the council", "the council", "get a panel", or wants multiple expert perspectives on a decision.
Forces the laziest solution that actually works, simplest, shortest, most minimal. Channels a senior dev who has seen everything: question whether the task needs to exist at all (YAGNI), reach for the standard library before custom code, native platform features before dependencies, one line before fifty. Supports intensity levels: lite, full (default), ultra. Use whenever the user says "ponytail", "be lazy", "lazy mode", "simplest solution", "minimal solution", "yagni", "do less", or "shortest path", and whenever they complain about over-engineering, bloat, boilerplate, or unnecessary dependencies.
Turns a vague ask into a rigorous, grounded, token-efficient prompt for an agent or LLM. Bakes in explicit role + objective, testable success criteria, anti-hallucination rules (verify or say "I don't know", never invent files/APIs/facts), anti-tokenmaxing (output budget, no restating the prompt, lead with the answer), and strict agent rules (read before edit, ask when blocked, small reversible changes). Use whenever the user says "prompt-generator", "write a prompt", "make a prompt for", "improve this prompt", "turn this into a prompt", "prompt for an agent", or hands over a rough/sloppy prompt to tighten.
Local, free, multi-specialist review of the working diff using parallel Claude subagents, then adversarial verification of every finding, dedup, and a ranked report with file:line evidence. Specialists: correctness, security/trust boundaries, data/perf (N+1, transactions), architecture-altitude, ponytail-simplicity, and tests/failure-paths. Complements the built-in /code-review (this is local + tailored; /code-review ultra is the cloud one). Use when the user says "review-swarm", "swarm review", "multi-agent review", or "deep review this diff".
Preflight sync + situational-awareness brief to run BEFORE starting work on a repo, so the agent builds on latest code, not stale local state. Fetches from origin, reports ahead/behind divergence, summarizes recent commits and open PRs/issues, flags uncommitted/dirty state, and (on request) refreshes library docs via context7. Read-only by default — never pulls or rewrites a dirty tree without explicit OK. Use when the user says "up-to-date", "sync first", "pull latest", "catch me up on the repo", or "before we start".