Context engineering

Keep the path, not the page

Lesson 0009 cleans a window that is already full. Just-in-time is how you avoid filling it: hold a cheap pointer, navigate, and load one page this turn — the rest of the kitchen stays on disk.

1. 1 · PointerA path, query, or URL. Cheap. The page stays on disk.
2. 2 · NavigateGlob is names. Grep is lines inside. Folder names are already signal.
3. 3 · LoadRead (or head/tail) this turn. Still propose · run · paste.

Same window as 0002. Same slips as 0005. The new move is what you hold before the slip.

The shelf is disk. The ticks in the bar are pointers. Only one page pays rent this turn.

You already met the cousins

RAG: your code finds a few pages and pastes them, then one call. A skill: you wrote a how-to folder; the catalog line is always in RAM; the chapter loads on a match. JIT: the kitchen already exists — a codebase, a drive, a database — and the agent walks it. Same essay: CLAUDE.md may drop in up front; glob and grep fetch the rest just-in-time.

Architecture first. RAG is pre-retrieve. JIT is load during the loop.

# STUFFED — disk copied into RAM
Read every file under kitchen/          # refunds, logs, dough…

# RAG — you already knew the two pages
paste(walnut.md, allergens.md)
ask("how do we make walnut crumb?")

# JIT — pointers in RAM, then slips
Glob("recipes/**")
# → recipes/walnut.md
#    recipes/margherita.md     # not loaded
Grep("walnut crumb", path="recipes/")
Read("recipes/walnut.md")               # this turn only
# oven/log.txt never entered RAM

Kitchen only to make the pointer concrete. The load-bearing picture is the bar and the two columns.

Glob is names. Grep is guts.

Claude Code: Glob finds files by pattern (recipes/**). Grep searches contents (built on ripgrep). Read takes a path and returns lines — and tells the model to page if the file is too big. You do not need those product names at work. You need the three rents: pointer, navigate, load. Cursor’s agent loop is the same shape. Same tools page · How Claude Code works

Skill: watch the same ticket shrink

Stuffed copies the shelf. RAG is you pasting two pages before the call. Pointers hold names. Navigate greps. Load reads one file.

Hybrid is allowed

Anthropic: Claude Code drops CLAUDE.md in up front (a cheap always-on note) and uses glob/grep for the rest. That is not stuffing the repo. Runtime exploration is slower than a pre-built index — use RAG when you already know the two pages; use JIT when the agent must discover them. “Do the simplest thing that works.” Same essay

Parked: sub-agents — a helper gets a clean window and returns a short summary, so the search dumps never hit your RAM. That is the other half of map item 10, not tonight. Prompt caching is still a cost trick, not a pointer.

Check