model-tiering

Model Tiering — route the work, keep the judgment

A strong model's scarcest resource mid-build is clean context. Every mechanical, fully-specified step it executes itself dilutes the attention that architecture, integration, and review need — so the plan's author grades each step for the cheapest tier that can execute it reliably, and execution honors those grades with subagents. The motivation is quality via focus, not cost. The failure modes are symmetric, and the bar in Section 3 resolves both: doubt about any criterion keeps the work up; a step that clears all four goes down — hoarding it spends the exact context this skill exists to protect.

How to apply this

Locate your tier first — before any step is tagged (Section 1). Then the skill fires at two moments:

• While authoring a plan — grade each step in place per Sections 2–4, tagging as you write. Don't defer tiering to a "who builds what" appendix: that retrofit framing is what produces all-or-nothing calls like "the session builds everything."
• While executing a plan whose steps carry tags — honor them per Section 5.

An orchestrating harness may carry the owner's standing disposition for a run — pre-authorize delegation, or suppress it entirely. Pre-authorization resolves borderline calls toward tagging among steps that already clear all four criteria of Section 3; it never relaxes those criteria, never overrides Section 4's categorical list, and cannot invent a tier that isn't there.

1. Know your position on the ladder

Identify the session model however the harness exposes it (in Claude Code the system prompt states it), find it on its own model family's ladder under Provenance, and orchestrate only if at least one trustworthy executor tier sits strictly below you. Otherwise note it in one line — "no executor tier below; executing inline" — and stop applying this skill: no tags, no subagents.

Never route upward, and never across families. The judgment this skill protects has to sit in the session — where the user's context and the approval gate are — not inside a stronger subagent that returns a summary; and a model whose family has no ladder under Provenance has no position, so it has no executor tiers. (Same-tier subagents exist only for [delegate-heavy] — Section 2.)