firstpasslisted
Install: claude install-skill undercutsh/firstpass
# Tiered Dispatch — confidence-gated model routing
Goal: every unit of work runs at the cheapest tier that can pass verification,
and escalation is driven by **objective signals** (failed checks, disagreement,
explicit uncertainty flags) — never by asking a model how confident it feels.
LLM self-reported confidence is poorly calibrated; this skill treats it as a
routing hint only, never as approval.
This skill defines the ladder logic and is model-agnostic. It uses four
generic tier names — **cheap**, **standard**, **frontier**, and an optional
**apex** — instead of any vendor's specific model names. Resolve tier names to
actual model IDs for your agent (Claude Code, Codex, Cursor, or other) via
`models.md` in this same directory.
## Scope: this rubric tiers execution, not judgment
A single request usually mixes two kinds of work: deciding what to build and
which tradeoffs matter (judgment), and then building it against that decision
(execution). This entire skill — the six-flag rubric, the escalation
triggers, the residue-only handoff — applies to the **execution** half. It
has no validated answer for the judgment half; treat planning/strategy calls
as frontier-tier by default until that's tested (see "Anti-patterns" and the
site's Honest limits, #05). Don't let "but the whole session touched a
frontier model at some point" become an excuse to skip tiering the execution
units that make up most of the token volume.
## The two routing decisions
1. **Planning tier** — who