Right-Size Every Claude Task with Token Squirrel 🐿️

Most people choose a Claude model once, leave it on the heaviest tier, and never touch the effort or research controls. That burns through the usage limit on work a lighter setup would handle just as well, and on simple, well-specified tasks a frontier model can overcomplicate the answer and return something worse. The controls that would fix this, model tier, effort, thinking, and research, are the ones most people do not understand well enough to set with intent.

Before starting a task, let the skill classify it on a five-level complexity scale, from an instant lookup up to open-ended frontier reasoning. Take its recommendation across four dials: model, effort, thinking, and research. When the current model is heavier than the task needs, read the usage barometer it shows, which compares where you are against where the task actually sits, then choose to proceed or switch down before any work begins. To switch, change the model, effort, or research in the picker and re-send, or start a fresh chat. Do not expect the skill to act until that choice is made.

The same quality at a fraction of the usage, with the decision made consciously once instead of by default. Teams stop reaching for the frontier tier on small tasks, stretch their usage limits further, and get better answers on the simple work where a heavy model actually hurts.

Keep every model name in one registry table and edit only that table when a model ships or is retired, since tasks map to durable roles rather than to names. Leave adaptive thinking on and treat effort as the real control. Let the skill check Anthropic's published model guide to confirm the registry is current. Expect a visual usage barometer where inline visuals render, and a one-line text prompt where they cannot, such as a terminal or some agent platforms.

From inside a conversation the skill can see the current model but usually not the current effort or research state, so those parts of its advice are recommendations to verify rather than readings of the actual settings. The skill cannot change any setting itself; every switch routes through the person's own controls. Usage cost is expressed as a rough relative magnitude, never an exact dollar figure, because the skill cannot see the account bill.