Token Discipline
Internal course · AI usage

Token Discipline

Everyone on the team is on Claude now. This is the difference between two people getting the same result — one for a fraction of the tokens. Short modules, a one-page cheat sheet, then a graded quiz.

5–10×Possible cost gap between two people getting the exact same result from Claude
30–80%Typical cut in spend from the habits in this course, with no drop in output quality
90% offDiscount Claude gets when it reuses context it just saw, instead of reading it again from scratch
0 / 6 modules
QUICK REFERENCE

Cheat sheet

A one-page summary now that you've been through the modules. Each line is explained in full above — use this as a refresher before the quiz.

Tokens & cost

Cost = tokens × model's per-token rate. The whole conversation resends every turn — a long stale session costs more even for a one-line question.

Model

Default Sonnet. Haiku for high-volume/simple. Opus/Fable only for genuinely hard reasoning.

Effort

Default medium/high. Drop to low for simple or high-volume work. Reserve xhigh/max for frontier problems.

MCP servers

Disable unused ones via /mcp. Prefer a plain CLI (gh, aws, gcloud) when one exists.

Context

Check /context before guessing. Move rarely-used CLAUDE.md instructions into skills.

Prompts

Name the file/line, not just the symptom. Cut "triple-check everything"-style rituals from standing prompts.

Long documents

Put the document(s) first, your question last — measurably better answers, fewer costly re-asks.

Structure

Wrap distinct content in XML tags (<context>, <instructions>) so Claude doesn't misparse and re-ask.

Search

Ask for a targeted lookup, not "explore the codebase" — vague asks trigger broad, expensive scanning.

Caching

Keep stable content first, dynamic content last. Don't touch the system prompt or tool list mid-session. Cache dies after 5 min–1 hr idle.

Sessions

/clear for new work, /compact mid-task. /rename before clearing. /rewind to undo a bad turn instead of arguing it back.

Subagents

Delegate verbose/exploratory work. Pin cheap ones to Haiku via /agents.

Task budgets

Give a long agentic task an advisory token budget so it self-regulates instead of running unbounded.

Agent teams

~7× the tokens of a normal session. Keep teams small, shut teammates down once done.

Projects (Desktop)

Put shared context in the Project once, not pasted per chat. Keep project knowledge lean.

Check yourself

Quiz — 18 questions

A self-check, not a test — answer each question, then see which ones you got right. Retake it as many times as you like.