Use when wiring the canonical cortex inbox+outbox polling loop into Claude Code's /loop. This is the orchestration spine — every empirica claude polls Cortex on a fast adaptive cadence (30s base, 5m max) for proposals addressed to itself + status changes on its own outgoing proposals. Self-throttles when an empirica transaction is open (the AI is already busy; no need to interrupt). The canonical loop catalog (empirica/core/cockpit/canonical_loops.py) auto-installs this when the TUI cockpit toggles L on an instance that has no loops registered. This skill is the body the AI runs each fire.
Daily housekeeping body for the canonical `message-cleanup` loop. Prunes expired git-notes mesh messages so the inbox stays focused on un-read ones. Loaded by the loop scheduler when the cron entry fires (default 03:17 daily) — never invoked directly by a user. Triggers: `<task-notification>` from the message-cleanup loop, "message housekeeping", "expired messages", "prune mesh".
Empirica Constitutional Decision Tree — the governance framework that routes situations to the right mechanism. Load this skill when unsure which Empirica mechanism to use, when starting a session, or when the system prompt feels insufficient. Replaces front-loaded instructions with a decision framework. Triggers: 'which mechanism', 'how should I handle', 'what tool for this', 'empirica constitution', 'decision tree', or any uncertainty about which Empirica feature applies to the current situation.
Use when starting complex work, planning implementation, breaking down tasks, creating specs, or when the user says 'plan this as transactions', 'plan transactions', 'break this down', 'create a spec', 'how should I approach this', 'transaction plan', or mentions needing a structured approach to multi-step work. This skill guides the full epistemic workflow from task decomposition through measured execution. Prefer this over EnterPlanMode for non-trivial tasks.
Use when sending a message to a PEER AI in the mesh — discussion, FYI, question, request to do work, or completion-ack for a request a peer made of YOU. Pairs with /cortex-mailbox-poll (the receive side). Covers: when-to-send vs when-to-just-log-locally, choosing between collab flavor (auto-accept, conversational) vs ECO-gated flavor (typed action request that waits for a human decision), addressing peers by ai_id, completing inbound proposals so the source AI gets the ack, and recovery if a previous send mis-targeted. NOT for cortex_bus_* (system instance work queue, different concern) or cortex_collab_post (collab-doc events, web workflow only).
Use when arming an event listener for the canonical mesh — when the user says 'arm this listener', 'subscribe to ntfy topic', 'wake me when X arrives', or when responding to a system-reminder from listener-install-pickup. The new canonical flow is `empirica listener on/arm/off` — three single-purpose tool calls that auto-resolve defaults, short-circuit when a persistent OS service is already subscribed, and emit structured next_step JSON the AI can mechanically chain. The older curl-based pattern lives as the 'legacy / custom topics' fallback at the bottom.
Use when scheduling the canonical biweekly services-audit cron loop for Empirica's AI service scanner (Phase 3). The auditor body is `empirica services-audit` — captures a fresh scan, diffs against the previous, and emits a notification when novel running services appear. This skill provides the prompt template that wires the body into Claude Code's `/loop` cron mode + Empirica's loop registry. Recommended cadence: `0 6 1,15 * *` (1st and 15th of each month at 06:00 UTC — biweekly).
Use when the user runs `empirica scan --explain` or asks you to audit running AI services. You read the deterministic scanner snapshot, judge each AI-touching process against the bundled security corpus, and emit findings/assumptions/unknowns with confidence + cited corpus sections. Two-tier judgment (cheap AI-touching pre-filter, then full taxonomy with citation). Read-only by design — never kill processes or modify configuration; emit `recommended_action` strings only. Tracks citation coverage explicitly (which of the 5 corpus files were referenced before each finding) so trust grounding is auditable.