Skip to content
BXGLNDR

PRJ-2026-17ProductOn request

Claude Code Setup, an environment tuned to cost less and break less

Every setting comes from a failure or a bill.

Project at a glance

The need
A Claude Code environment tuned from day one: lower usage, guardrails and an installation that checks itself.
My contribution
Designed, wrote and tested the package: rules, settings, hooks, sub-agents, measurement tools and install scripts, drawn from months of daily use.

What was built

  • Rules and measured settings
  • Four hooks
  • Three sub-agents
  • Install and check scripts

Version 4 of 25 October 2026, installed and checked on my own machine. Available on request; no sale or third-party installation is claimed.

This file is text: no screenshot, on purpose.
Nature
Installation package for Claude Code: rules, settings, hooks, sub-agents and scripts
Period
Since June 2026
Client
Setup Claude CodeIn-house product, available on request

The need

Claude Code installs in one command, but daily use calls for settings that no documentation gathers in one place: which model for which task, when to compact the context, how to stop a command from reading a secrets file, how to tell that a setting has stopped working. Without them, usage climbs and failures go unnoticed.

What the package contains

  • Global rules for Claude: the cheapest model that will do each task well, short sessions closed on a checkpoint, nothing irreversible without your approval.
  • Measured settings: the context window size and the compaction threshold come from replaying 45 days of real sessions.
  • Four hooks: compacted command output, protected secrets files, public web pages read by a remote browser rather than a full one, a warning when a rules file grows too heavy.
  • Three sub-agents, one per model tier: search and read, apply a decided change, think through hard cases without touching anything.
  • Four tools to measure and cut usage: a breakdown by day, model and project, the cache share, a compaction replay, and a splitter for a rules file that has grown too heavy.
  • Two scripts: one installs, backing up every file it replaces first; the other checks the installation end to end.

Every setting comes from a failure or a bill

A rules file of nearly 600 KB made every session start at about 290,000 tokens instead of 70,000. A badly configured background plugin took up to a third of total usage. A hook left unwired compacted nothing for over a month, and nothing broke. The package turns these into rules, and its check script actually runs every hook with a test input to catch such silent failures. The lessons ship with it: you know why each setting is there, and what you can change.

Evidence

Every figure carries its source. No commercial result is claimed.

  1. 70 %

    Approximate share of usage made of cache reads, on real use

    SourceMeasured with the cache-audit tool, the package’s lessons, 25 October 2026

  2. −52 %

    Replayed cost over 45 days of sessions, 300k window versus no limit

    SourceReplayed with the cap-replay tool at Opus pricing, the package’s lessons, 25 October 2026

  3. 290k → 70k

    Tokens at session start, with and then without a 600 KB rules file

    SourceThe package’s lessons, 25 October 2026

  4. 4

    Hooks run with a test input at every check

    SourceThe package’s check script, version 4 of 25 October 2026

What it does not contain

  • No secrets: no keys, no tokens, no credentials. The MCP server examples carry placeholders to fill in.
  • No skills by other authors: they are downloaded from their original repository, with their licence.
  • No machine-specific paths: the installer puts in your own folder.

Who it is for, and how to get it

For a developer or a small team using Claude Code on Windows 10 or 11 who wants a tuned environment from day one; on macOS or Linux, each file is installed by hand. Installation takes about twenty minutes. The package is available on request: write to me and I will reply with the terms.

Technical details

The technologies used for this project. Features and their limits are detailed in the narrative.

  • Claude Code
  • PowerShell
  • Node.js
  • Markdown

Expertise involved

A similar need?

Let's discuss your context and what you would like to simplify.

Talk about your project