LEARN TO USE AI BETTER AND CHEAPER

THE 8 GAMES

• Cache Me If You Can: survive 30 days on one AI budget by catching cache hits and dodging runaway agents.
• Model Match: pick the cheapest option that does each job well, from no AI to Fable 5.1, at real prices.
• Dial It In: choose when a job runs (now or batched at half price) and how hard the model should think.
• Context Packer: give the AI only the context a task needs, and keep secrets out.
• Prompt Smith: sharpen vague prompts with audience, goal, format, limits and examples.
• Trust, but Verify: flag the claims in an AI answer you must check, and learn how to check them.
• Loop Breaker: rescue chats, agents and rollouts that go sideways.
• Prefix Stacker: order prompt blocks so prompt caching can reuse them, and watch the bill drop.

LEARN AI: PLAIN-ENGLISH COURSE

New to AI? 9 short chapters (about 52 minutes) for people who do not write code, each with a quiz. Chapter 1 is free; the full course is a one-time $2.99. See the course or take it interactively.

1. What AI is: What tools like Claude actually are, what they are great at, and where they trip up.
2. How it works: Learning, tokens, why answers change, and why long chats get heavy.
3. Asking well: How to brief Claude so the first answer is close to what you need.
4. Giving it your material: Files, Projects and memory: how to give Claude the right background without overloading it.
5. Search, Artifacts, Connectors and Skills: The add-ons that let Claude look things up, build things, reach your apps and follow your playbooks.
6. Checking its work: What to double-check, how to check it quickly, and why you stay responsible.
7. Safe, private and fair: What never goes into a chat, what happens to your chats, hidden instructions, approving actions, and fairness to people.
8. Using it well, within your limits: Models, effort, usage limits and when not to use AI at all.
9. Putting it all together: One real task from start to finish, how to find your own uses, and how to keep getting better.

THE 16 HABITS YOU WILL LEARN

1. Right-size the model. Use the smallest model that does the job well. Save the strongest models for work that is long, subtle or expensive to get wrong. The saving matters most at volume.
2. Not everything needs AI. Exact, repeatable jobs (sums, format conversions, validation, searching files) are cheaper and safer as a formula, command or script. Ask AI to write the script once.
3. Count cost per finished task. A cheap model that needs three retries is not cheap, and a pricier one is only worth it if it finishes more. Judge the cost of getting the job done, measured on your own work.
4. Set effort, not magic words. How hard the model thinks is a setting (low, medium, high, xhigh, max). Words like "think hard" are ordinary text that only nudges. Defaults differ: medium in Claude Code, high for Sonnet 5.5 on the API.
5. Batch what can wait. Bulk work nobody is waiting on goes through the Batch API at half price. Results can take up to 24 hours and each request is single-shot, with no tool loop.
6. Give the right context. Give what the task needs and nothing else. Size is not the test: a big document can be essential and a small one useless. Point an agent at a file instead of pasting a folder.
7. Keep secrets out. Passwords, keys, card numbers and other people's personal data never go into a prompt unless the task truly needs them and you are allowed to share them. Treat text an agent fetches as data, not orders.
8. Start fresh between tasks. Old context keeps getting sent, clutters the model's attention and uses up limits. New topic, new chat (/clear in Claude Code). Corrected the same thing twice? Start over with a better prompt.
9. Stable first, changing last. Put content that never changes (instructions, tools, documents) first and changing bits (time, IDs, the question) last, so the prompt cache can reuse it at a fraction of the price. Slow traffic needs the 1-hour cache.
10. Be specific. Say who it is for, what good looks like, the format and the limits. Ask for short, structured output when code will read it: output tokens cost about 5x input.
11. Show, and say why. Show an example of what you want and explain why a rule exists. Both beat shouting rules in capitals.
12. Check what would hurt. Check exact numbers, quotes, sources, links, package names, versions and legal, medical or security claims before you rely on them. Let the AI say "I don't know", and test a new model on your own tasks.
13. Give it a check to run. "Looks done" is not done. Give the AI a test, typecheck or other pass/fail check so it can prove the work.
14. Put agents on a leash. Before an agent runs alone, set a turn limit, a spend cap and a test that means "done". Gate destructive commands, roll out with a pilot, and make spend visible per team.
15. Plan before big changes. For big changes, plan first and review the plan. Stop early (Esc) and rewind when it heads the wrong way.
16. Keep your setup lean. What loads every time costs every time: keep CLAUDE.md short, disconnect unused tools, use skills on demand and subagents for noisy work.