Every turn re-sends the whole conversation. This course shows where the tokens go and how to cut them: cache-friendly prefixes, cheaper models, /clear and /compact, subagents and compression tools.
A token map of your own CLAUDE.md, checked against /context
A restructured CLAUDE.md cascade with global, project and gitignored local layers
Return contracts and Haiku subagent dispatches for research tasks on your codebase
A working caveman + rtk setup with your own measured baseline
A baseline-vs-tuned run of one real workflow, comparing tokens, time and turns
Tokens vs words, the input/output split, what prompt caching does and does not cache, the context window, and the resend tax that makes long sessions expensive.
Per-model pricing for Opus 4.7, Sonnet 4.6, and Haiku 4.5 — input/output rates, cache rates, quota weight on Pro and Max plans, and a decision matrix for picking the right model per task.
When to /clear, when to /compact, when to fork-session for risky experiments, and how to recognize context rot before it doubles your bill.
Trim CLAUDE.md, layer with @includes and CLAUDE.local.md, exploit prompt caching by keeping the prefix stable, and write memory entries that the selector actually surfaces.
Specificity over verbosity, file references over context dumps, ask-once patterns, structured output contracts, and the four anti-patterns that quietly inflate every turn.
The 200-word return contract, when to offload vs do inline, parallel dispatch, and background agents that survive independently of your main conversation.
Two open-source tools that compound to ~80% savings: caveman (output compression Claude Code plugin) and rtk (tool-input compression via Bash hook). Install, daily use, before/after measurements.
Pro vs Max vs API+caching decision tree, a monthly budget audit checklist, and a capstone exercise that walks through tuning a real session from baseline to 80%-cheaper.
One year of access for $20, or subscribe for $45/month to get this course and every other one.
Claude Code and real sessions to inspect. Every hands-on runs on your own sessions and your own CLAUDE.md, not a toy repo. It works on Pro, Max or the API.
8 tutorials, each with hands-on steps, exercises and a quiz. The listed lesson times add up to about 3.5 to 5 hours. Some exercises run over several days on purpose, such as five days of /clear.
Two open-source tools in the compression tutorial: caveman (a Claude Code plugin that compresses model output) and rtk (a CLI proxy that compresses shell command output). Both are optional. The rest of the course uses built-in Claude Code commands only.
No. It stands on its own. If you know Claude Code basics from Fundamentals, you will move faster. The habits apply to any Claude Code work, solo or team.