The Bill Guard

# The Bill Guard

Six rules decide which brain runs every job you hand your AI. Right now, none of them are yours.

Renaming a file. Sorting a list. Formatting some json. None of it needs a genius, but nothing in your setup asks. The model grabs the biggest, priciest brain for every task, every time, and quietly bills the whole trivial 90% at premium rates. The leak isn't that your AI is expensive. It's that the decision about which brain runs the job is being made for you, by default, and you find out on the invoice.

The Bill Guard is those six rules, written down and countable, so the decision becomes yours: the cheap 90% goes to a small model, and the expensive brain is reserved for real judgment. Same jobs. One rule picks who runs each one.

Three of them are the tier rules. You map CHEAP / MID / TOP onto whatever three models you already have, once, and each one comes with the question that routes to it. (The kit states these twice: once as a table of tiers, once as a one-line decision. They are the same three rules from two directions, which is exactly why this is six rules and not nine.)

• **CHEAP** runs the mechanical majority: reading, grepping, mapping, classifying, grading, summarizing, formatting, renaming, sorting, extraction. *Ask: does this job just need something read, sorted, classified, or checked?*
• **MID** is the default *builder*: code, refactors, deploys, repo-reading, research, most execution. With a good brief, MID executing beats TOP typing. *Ask: does this job need something built or changed?*
• **TOP** is judgment only: multi-stage planning, red-teaming, distilling rules from evidence, grading a finished plan, the final verdict. It produces plans and verdicts, not volume. *Ask: does this job need a judgment call or a plan?*

Two are the guard-rules that stop silent bill creep, and this is where most of the savings live:

• Cheap by default, expensive on purpose. The default is CHEAP. Reaching for TOP is an explicit, deliberate choice, never the silent fallback. The leak is always "everything quietly ran on the expensive model."
• Sub-tasks route DOWN, never up. Any helper or sub-agent a big job spawns is MID-or-lower, and read-only helpers are CHEAP. A verifier *should* be cheap: a fresh, cheap, independent context is exactly what checking needs. Always pass the model choice explicitly; "inherit the parent's model" is how the whole tree ends up on TOP.

And the sixth is the anti-pattern to grep your own setup for: *the top model typing what the mid model could type.* If your most expensive model is doing mechanical work, that's the leak.

It will not tell you when the cheap model got something wrong. There is no confidence score in here and no automatic escalation — the rules route by the *kind* of job, not by how well it went. Catching a bad cheap answer is still your call, and ROUTING-RULES.md says so out loud rather than pretending otherwise.

• `ROUTING-RULES.md` contains the six rules above, in full, tier-agnostic.
• `cheap-default.config.md` — a drop-in block for your system prompt / CLAUDE.md / agent config that makes the cheap model the default and forces an explicit override to reach the expensive one.
• `routed-job-example.md` walks one real job routed end to end, so you can watch the rule decide.
• `SETUP.md` lists the three steps to wire it in.

It's tier-agnostic. Map CHEAP / MID / TOP onto whatever you have: Claude Opus/Sonnet/Haiku, GPT-5/mini, Gemini Pro/Flash. The names don't matter; the tiers do.

1. Map your three models cheapest → priciest and label them CHEAP / MID / TOP.
2. Paste the cheap-default block into your system prompt / CLAUDE.md, tier names filled in. This alone plugs the biggest leak.
3. Route sub-work down: helpers are mid-or-lower, read-only checkers are cheap, and always pass the model choice explicitly — never inherit.

Then grep your own setup for the anti-pattern. If your priciest model is reading, sorting, or classifying, route that step down.

## Get it

Comment BILL on the reel and I'll send it over, or just grab it from the button above, no comment needed. Nothing expires and nothing is held back: the same six rules, the cheap-default config, and one job routed start to finish. Pay the setup once; the trivial 90% never touches the expensive brain again.

_Built by SharpenDaily — your AI, sharpened._