← ClaudeAtlas

confidence-gated-routinglisted

Use before delegating work to multiple agents or model tiers; assign the cheapest verifiable tier, escalate only objective failure or disagreement residue, and preserve an auditable handoff.
sandbaseai/workbuddy-skill · ★ 9 · AI & Automation · score 80
Install: claude install-skill sandbaseai/workbuddy-skill
# Confidence-Gated Task Routing Use this skill before fan-out, swarm, parallel-agent, or model-tier dispatch. The goal is to run each unit at the cheapest tier that can pass independent verification. Do not route by asking a model for a numeric confidence score; self-reported confidence is a hint, not approval. ## Separate planning from execution Choose a planning tier once, based on ambiguity and cross-cutting judgment. Choose an execution tier separately for each planned unit. Keep easy, mechanically verifiable work cheap even when the overall project is difficult. One hard unit must not raise the tier for the entire batch. ## Assign a base tier with six flags For each unit, mark: - `UNVERIFIABLE`: output cannot be checked by tests, schema, execution, or an independent sample; - `AMBIGUOUS`: multiple defensible answers require judgment; - `BLAST`: work is irreversible or touches money, authorization, production, deletion, or sensitive data; - `CROSS-CUTTING`: reasoning spans many files, systems, or contracts; - `NOVEL`: the task needs new design rather than an established pattern; - `FORMAT-STRICT`: output must satisfy an exact schema or protocol. Start with `cheap` for zero flags, `standard` for one or two, and `frontier` for three or more or any ownership/judgment call. Resolve these generic names through the host's explicit model map. `apex` is reserved for a sentence explaining why its marginal judgment value justifies the cost. When output is mechanically verif