AI Agent & Automation Workflows

Decision protocol for consulting Claude Fable 5 (the top-tier, usage-billed model) as a strategic advisor from an Opus orchestrator session. Use BEFORE committing to any decision that would be costly to revert — system or workflow architecture, database schema and data-model design, API/webhook contracts, n8n workflow topology, technology and vendor selection, production migrations — even when you could decide yourself; deciding solo on a one-way door is exactly the failure mode this skill prevents. Fires on phrasings like "make the call", "decide between X and Y", "we'll live with this choice", "one-way door", "before we freeze it", "has to be right". Also use BEFORE starting any unattended loop, schedule, or routine — /loop, /schedule, /goal with a high turn cap, cron-style agents, proactive workflows — because a loop design flaw repeats every iteration; Fable reviews stop conditions, interval, verification, blast radius, and cost per iteration once, before it runs. Also use when stuck after 2+ genuinely different failed fix attempts, when finalizing an implementation plan (including in plan mode, before ExitPlanMode), before declaring done on hard-to-revert work (production deploys, migrations, client-facing deliverables), and whenever the user mentions Fable, a second opinion, an advisor, or checking with a stronger model. Covers when a consult is worth 2× Opus token prices, the compact briefing packet, per-task budget caps, and how to weigh the advice.

Delegation protocol for sessions running on Claude Fable 5 (or any premium model) — plan big, execute small. Use BEFORE pulling any bulk material into your own context — sweeping a codebase, triaging logs, reviewing a document set, researching across many web pages, verifying N facts against sources, auditing every workflow/config/endpoint, or any task where the reading is mandatory and voluminous. Critical because subagents inherit the session model — on a Fable session an un-pinned Explore or general-purpose spawn bills at Fable rates, so delegation alone saves nothing; this skill makes the rate split real with parallel workers pinned to Sonnet/Haiku that read in their own contexts and report distilled findings (~2.5× cheaper, ~3× faster, 84–98% of input at worker rates in Anthropic's cookbook measurements). Fires on phrasings like "sweep", "audit all", "go through every", "check each", "triage these logs", "research across", "verify against the docs", "keep the cost down", "fan out", "delegate the reading", "plan big execute small". Covers the delegate-or-read gate, the worker brief format, model-tier choice, brief granularity, premise verification, and when NOT to split (narrow reads, judgment that needs frontier eyes on the raw material).