Cheat sheet

Context window

Compressed facts from lesson 0002. Print this after the labs, not instead of them.

What it is

• The maximum token sequence the model conditions on in one call. Karpathy: RAM. Anthropic: attention budget.
• Shared by system, history, tools/files, and the output you still need to generate. Overflow → API error.
• Counted in that model’s tokenizer, not words. Token cheat sheet.

Two failure modes

• Hard: over the cap. Rejected.
• Soft: under the cap, but context rot (recall fades as n grows) and the U-curve (middle facts weaker than start/end). Anthropic; Liu et al.

A receipt, not a vibe

• OpenAI cookbook: "tiktoken is great!" → 6 tokens on cl100k_base (IDs 83, 8251, 2488, 382, 2212, 0).
• Same word, different encodings: antidisestablishmentarianism is 5 tokens on r50k/p50k and 6 on cl100k/o200k.
• Chat messages add wrapping tokens; max_tokens reserves output inside the same window.

Why later patterns exist

• Smallest high-signal token set — the seed of context engineering.
• RAG, skills, evict · compact · note, sub-agents: ways to page RAM to disk instead of stuffing the window. Context engineering card.