The setup and content

The best out of Claude, for the least cost. Eight parts, one cost lever each.

Part 1 of 8

What every message costs

Read your usage: watch context per step and step count.

Copy this Estimate context times steps, then write down the measured number.

Part 2 of 8

Rules that change behaviour; cut the rest.

Copy this Keep a rule only if it changes behaviour.

Part 3 of 8

The smallest model that does each step well.

Demo

My model policy
Who or what	Model
Talkers: the project bosses	Sonnet 5.5, high
Workers, normal project work	Sonnet 5.5
Workers, small mechanical steps	Haiku 5.5
Research and planning	Opus 5.5, high
Chat plumbing	Haiku 5.5

One task, three steps, three models.

Copy this Start small; move up when a wrong answer costs trust.

Part 4 of 8

Heavy reading goes to helpers; the chat stays light.

Lever: The main chat is re-read on every step.
From my setup: Foreground helpers once blocked my chat for 22 of 34 minutes.
Avoid: Big logs in the main chat; compaction wiping what you loaded.

Copy this Read big things in a helper; keep the summary.

Part 5 of 8

Pick the cheapest worker that works.

Lever: Scripts cost nothing; narrow helpers beat one giant agent.
From my setup: Middle model, 3 items a session, short playbooks: a tenth the cost.
Avoid: One giant agent holding every file.

Copy this Script first, one narrow helper next, workflows for many items.

Part 6 of 8

Write it down once; search it later.

Copy this Decide once, with a key; search before you ask.

Part 7 of 8

Load each tool only where it is used.

Copy this Connect a tool where it is used; outward actions need a yes.

Part 8 of 8

Leave a job running; come back to a result.

Lever: Idle hours and silent stalls cost too.
From my setup: A prompt after I left blocked my chat overnight, until 11:53.
Avoid: A prompt that waits for a human.

Copy this Before you leave, ask every permission once and launch first.