AI Agent Workflow Collection

Tiered model routing — classifies incoming tasks by complexity and dispatches to the correct model tier: Tier C (Haiku, simple/high-volume), Tier B (Sonnet, mid-complexity), Tier A (Sonnet + Opus advisor, strategic/complex), Tier A+ (Opus solo, highest stakes). Uses Anthropic's native advisor20260301 API tool for Tier A dispatch. Enforces plan-before-act discipline: Opus advises on the PLAN, executor carries out the plan. Budget control via maxuses parameter. Trigger on: "what tier is this", "classify this task", "dispatch this", "which model should handle", "advisor mode", "route this task", "run this as tier A", "check my advisor budget", "run with opus advisor", "advisor dispatch", "tier routing".

Use when users want to set up or connect to the Atelier Brain -- the persistent memory layer for the pipeline. Auto-fixes existing config silently; guides new users through first-time setup.

CEO-level cross-department briefing for {{operatorname}} at {{companyname}}. Reads all department Brain sessions and decisions, synthesizes into a structured brief covering what moved, what's stuck, pending decisions, and upcoming deadlines. Use this skill whenever {{operatorname}} says: "brief me", "what's the state of the company", "morning brief", "company status", "what happened", "catch me up", "how are we doing", "daily brief", "what do I need to know", or at the end of any sessionend to close the loop. Also triggers on: "what's moving", "macro view", "give me a company update". This is the Chief of Staff brain — always run it when the operator needs a company-wide picture, even if they don't say "brief" explicitly. NOTE: backlog-lane intents (aging / pending / unresolved decisions) route to cross-dept-decisions, not here. See {brainroot}/AGENTS.md §Known DRY pairs.

Meta-audit of the skill system. Walks {brainroot}/skills/ + {brainroot}/AGENTS.md and reports three classes of drift: (a) skills with no resolver row (unreachable), (b) AGENTS.md rows pointing at missing skills (orphan phrases), (c) DRY overlap — two skills claiming the same trigger substring. Exit code 0 = clean; non-zero = drift. Trigger on: "cerebro doctor", "check resolvable", "skill health check", "resolver audit", "DRY audit", "are my skills reachable".

Adversarial parallel comparison — runs the SAME task across multiple model tiers and/or providers in parallel, then surfaces where each model disagrees, fails, or hallucinates. Purpose is NOT consensus: the purpose is exposing failure modes so the operator can route around them. Built on top of advisor-mode. Trigger on: "assemble the council", "convene the council", "put it to the council", "what does the council say", "council mode", "run this on all tiers", "compare model outputs", "which model failed on this", "adversarial compare", "diff the models", "run the council", "parallel dispatch", "show me the disagreements".

Validate and auto-inject the 11-key provenance frontmatter contract on memory files at write time. Use when authoring memory entries, running lint passes on memory directories, or backfilling missing attribution fields. Enforces a uniform provenance schema (name, description, type, sourcesurface, host, tool, sessionid, writtenatutc, sourcepath, confidence, lifecycle) so every memory entry is traceable to its origin surface, tool, and session.

End-of-session ritual — reflects on what happened, stores a structured session summary to the correct department Brain, updates the Master Dashboard, updates the Master Brain awareness layer, and updates local memory. Use when the user types /sessionend or says they're done for the session, wrapping up, or signing off.

Session opening ritual v2.4 — adds Stage 0 session-continuity probe (short-circuits full brief on same-day reopens within 1h break window), single-shot Stage 2 brief loader (one Bash call surfaces dept CLAUDE.md, latest session, pending decisions, dashboard header, memory index, and infrastructure state), runs ground-truth verification against filesystem (including a recursion guard so corrections themselves verify before firing), auto-corrects stale memory, delivers a Department Head Brief plus Infrastructure Snapshot, captures session intent, and renders the same brief to an HTML archive. Use when the user types /sessionstart or says they're getting started, picking up where they left off, or starting a new session.