Free · No signup · Beginner to expert

Use AI like
an engineer,
not a search box.

Everything that makes AI useful comes down to one thing — what you put inside the context window, and what you leave out. More context does not always mean better answers.
The model may process everything you send, but not everything receives equal attention.
The skill is not giving AI more information.

Red is the problem · Green is the fix · Blue is just a fact

0%

of a typical agent payload is sent every turn and never read

0%

cheaper on cached reads once your prefix is ordered properly

0

to index a skill the agent might need, instead of thousands to load it

0

memory a model has of yesterday — everything it knows, you put there

See it once and it sticks

One task. Two architectures.

The same request — "refactor this PDF parser" — packed two ways into the same 200,000-token window. Press the other button and watch the red disappear.

Context window · 200,000 tokens

Window filled —

Never read —

Cost per turn —

0100k200k tokens

    New here? What am I actually looking at?

    AI models read text in chunks called tokens — roughly three quarters of a word each. Everything the model can see at once has to fit in a fixed space called the context window, and you pay for every token you put in it, every single time you send a message. The red stripes are text that was sent and paid for on every turn without ever being useful. Phase 1 explains this slowly; Phase 2 shows you how to fix it.

    Illustrative figures, priced at $3 per million input tokens. Token counts are modelled, not measured — Phase 2 shows you how to measure your own.

    Why AI keeps forgetting

    A model has no memory of its own. Everything it appears to remember, someone put back in front of it.

    Every message starts it from nothing — no recollection of yesterday's decision, no idea what you agreed an hour ago. That is true of ChatGPT, Claude, and Gemini alike; none of them has quietly fixed it. Which means the real skill was never phrasing a clever question. It is deciding what gets loaded, in what order, and what gets dropped first when the room runs out.

    Learn how to fix it →

    Four phases.
    Zero to agents.

    Ethical use of AI

    Getting it right matters more than getting it fast.

    What to verify, what never to paste in, and where the line sits between assistance and misconduct. Separate guidance for everyday work and for academic study.

    Everything is a file.

    The book comes with a repository, and the repository is half the point. Skills, memory templates, prompts, and evaluations are plain Markdown you can read, copy, and drop straight into a project of your own.

    Nothing here is a screenshot of a technique. If a chapter says a structure saves tokens, the structure is in the tree and you can run it yourself.

    contextbook/

    skills/        reusable instructions, one folder each
    prompts/       templates worth stealing
    memory/        MEMORY.md, TASKS.md and friends
    workflows/     end-to-end builds
    agents/        loops, planners, tool sets
    mcp/           servers and configs
    evaluation/    how you know it got better
    datasets/      small, honest test sets
    research/      sources, tiered by trust
    

    Written in the open, corrected in the open.

    This field moves faster than one person can check. Every page carries its sources, marked by how much they can be trusted — official documentation, peer-reviewed work, and a company describing its own product are three different things. Find a wrong number and it gets fixed in public.