A Claude Code skill · field tool Nº 001

Scrooge

The Claude Code skill that refuses to overpay. It routes every piece of AI work to the cheapest model that can do it right: frontier brains where it counts, penny labor everywhere else, and a receipt at the end showing what you kept. Type /scrooge. That's the whole workflow.

new to all this? see how it works, in plain english →

Nº 01 · the problem

Every model bills
like a partner.
Most of the work
is paralegal.

Hand Claude a big task and one expensive model does all of it. The same frontier brain that architects your system also greps your files, renames your variables, and reads your logs.

That grunt work is ninety percent of most jobs, and a model a tenth the price does it identically. You'd never staff a project this way. Your AI shouldn't either.

what a million tokens costs, by model

blended per-million-token rates, claude api list pricing, typical 70/30 input/output split. the gold bar is where scrooge sends the grunt work.

Nº 02 · how it works

A foreman,
not a freelancer.

Scrooge turns your session into a job site. Your main model stops doing the labor and starts running the crew. Four moves, every task, automatically.

Move 01

Decompose

The task becomes a graph. Every node gets classified: mechanical, judgment, or genuinely hard. No guessing, no routing by mood. Stakes never inflate a node. Only difficulty does.

Move 02

Route

Mechanical work goes to the cheapest capable model. Bounded judgment goes mid-tier. The truly hard calls go to the strongest model available, at full effort. It reads your setup first, so it works whether you run Haiku or Fable 5.

Move 03

Swarm

Independent work runs in parallel, in waves, the moment it unblocks. Every agent gets a precise brief and a hard output cap, so a dozen workers never flood the room with essays.

Move 04

Verify, then bill

No agent's word is taken for it. Artifacts get checked, failures escalate one tier with evidence attached, and the run closes with a receipt: what it would have cost, what it did, what you kept.

Newer still, Scrooge decides who runs the crew. When the work is routine, it lets a cheap-to-run coordinator manage the busywork instead of your top model, and it hands control straight back the moment a hard call shows up. That quietly kills the biggest hidden cost, the manager re-reading the whole conversation on every turn, without ever putting your quality at risk.

quality is the constraint scrooge optimizes under, not against. cheap output that fails a check gets escalated, never shipped.

Nº 03 · the math

Napkin math,
real prices.

Nothing here is a growth-hack number. It's Claude API list pricing and the boring arithmetic of sending cheap work to cheap models.

10× Price gap between the frontier model doing your greps today and the cheap one that could.
~70% Share of a typical big task that is mechanical: reads, greps, renames, tests, surveys.
60–85% Typical estimated spend cut on large tasks once routing does its job. Your mix will vary.
$10 Once. No subscription, no seat license, no upsell waiting behind the door.

worked example · one real-shaped task

The 3-million-token refactor.

A serious codebase job: survey, plan, edit, test, verify. Run straight through Opus, all three million tokens bill at frontier rates.

Scrooged, roughly 80 percent rides the cheap tier, 15 percent gets mid-tier judgment, and the hardest 5 percent still gets the frontier brain it deserves. Same output. A third the invoice. On a Fable 5 session the gap roughly doubles.

estimated cost · same task, two ways

estimate at list pricing: 3M tokens at ~$11/MTok blended, vs 80/15/5 routed across haiku, sonnet, and opus. run it daily and the difference is thousands a year.

On a metered API plan, Scrooge pays for itself the first serious afternoon. On a subscription, it stretches the same usage limits across three to five times the work. Either way, the receipt does the bragging. the counting-house arithmetic

Nº 04 · the bit

The Dickens
is rationed.

Scrooge has a personality: a reformed Victorian miser who grumbles once when the work starts and files one mock loss report when it lands. Two lines of theater per task. That's the entire ration.

In between, silence and work. No jokes mid-task, no bits over broken builds. The receipt uses your run's real delegation counts and real list pricing, clearly labeled as estimates. The comedy is small. The number at the bottom isn't.

Claim Your Receipt humbug included · free of charge
📉 Scrooge receiptcodebase audit
scouts through the fog11
drones on the edits6
auditors, double-checking4
frontier, doled out like pudding
without scrooge$41.80
with scrooge$11.20
you kept$30.60 (73%)

Somewhere in San Francisco, a ledger weeps. Estimates, not an invoice. But the ledger doesn't lie by much.

Nº 05 · in the box

One file.
Every session,
cheaper.

  • The Scrooge skill, installed with one unzip. Type /scrooge and hand it any task.
  • Model-agnostic routing. It reads your session first, then adapts: Haiku base, Sonnet base, Opus, Fable 5, all covered.
  • Works with vanilla Claude Code or a custom agent roster. No dependencies, no config, no API keys of its own.
  • Quality guardrails baked in: artifact verification, two-strike escalation, and a hard rule that cheap output never ships unchecked.
  • The receipt. Real delegation counts, real list pricing, honest estimates of what you kept.
  • The desktop meter for Mac: a live paper-receipt meter in your menu bar, a global hotkey, a lifetime ledger with history, drag-to-place, all local.
  • Every-outlet setup, built into the installer: a guided walkthrough adds Scrooge to the Claude app, chat and Cowork, and to claude.ai on the web. Three steps, about a minute, on by itself the moment it uploads.
  • No telemetry: your ledger never leaves your Mac.

It's a skill, not a platform. Nothing phones home, nothing subscribes you to anything, nothing sits between you and your work.

You get a precisely written instruction set that turns any Claude Code session into a disciplined, parallel, cost-routed operation. Built and battle-tested by Paul Von Rieter across the builds on the operator site: tEE, CRUMB, WERKS, GainSail, and the rest of the arsenal.

the offer

Ten dollars.
It tips itself.

  • Instant access: checkout lands you on the install page
  • Installed in about sixty seconds, one unzip
  • Works on every Claude Code plan, API or subscription
  • Yours forever. No renewal, ever.

field tool Nº 001 · digital delivery

Scrooge


claude code skill · lifetime license $10
Buy Scrooge

secure card checkout by stripe. you'll be redirected to the install page the moment payment clears.

Includes the Scrooge skill and the Scrooge desktop meter for Mac. One purchase, both tools.

Nº 06 · fair questions

Asked, answered.

What exactly am I buying?
A Claude Code skill: a precisely engineered instruction file that installs into your skills folder in one unzip. From then on, /scrooge turns any task into a routed, parallel, verified operation with a savings receipt at the end. No software to run, no account to create.
Which models and plans does it work with?
All of them. Scrooge reads your session before it routes: it knows whether it's sitting on Haiku, Sonnet, Opus, or Fable 5, and it adapts the crew accordingly. API accounts see the savings in dollars. Pro and Max subscribers see it as usage limits that suddenly stretch several times further.
Does it work in the Claude app or on claude.ai?
Yes, everywhere Claude lives. Claude Code is home base: the terminal and the Claude Code desktop app for Mac share one install, full crew, full meter. And the installer now makes sure Scrooge lands on every other Claude outlet too: on first install it walks you through adding the skill to the Claude app, standard chat and Cowork, and to claude.ai in your browser. Three guided steps, about a minute, and the skill switches itself on the moment it uploads. One honest note: those chat surfaces load Scrooge as its playbook, while the subagent crew and the menu-bar meter do their work in Claude Code. Claude Code on the web is covered per project by committing the skill into the repo's .claude/skills folder.
Won't cheap models make the work worse?
That's the failure Scrooge is engineered against. Cheap models only get work with a clear spec and a checkable result. Anything needing real judgment routes upward, every artifact gets verified, and a task that fails twice at a tier escalates automatically. The hard thinking never leaves the strong model.
Are the savings numbers real?
They're honest estimates built on Claude API list pricing, and the receipt says so on its face. Exact savings depend on how much of your task is mechanical. Big jobs with lots of reading, searching, and editing save the most. A one-line question saves nothing, and Scrooge won't pretend otherwise.
How do I install it?
Checkout redirects you to a private install page: download the ZIP, unzip it into your Claude skills folder with the copy-paste command provided, open a fresh session, and type /scrooge. About sixty seconds, and the page walks you through every step.
What if it's not for me?
Email info@theeditorialedit.com and a human will make it right. It's ten dollars. Nobody is fighting you over it.
Does the meter phone home?
No. The desktop meter runs locally on your Mac and keeps its ledger there. No telemetry, no account, no server it reports to. What it tracks stays on your machine.
Get Scrooge · $10 still reading? the receipt is more fun than the faq.