Central Info Station — AI skill for Claude Code

Central Info Station (CIS)

**Measured on a real multi-agent project:**

	before	with CIS	saving
context at session start	~50 KB	≤ 4 KB	~92%
one lookup ("where did we decide X?")	a whole doc, 50–300 KB	≤ 3 KB	94–99%
cost of the read gate per tool call	—	~40 ms	—

**Retrieve, don't read.** A self-updating index of your project's decisions, plans, handoff notes and agent rules — plus Claude Code hooks that make AI coding agents *use* it instead of pouring 300KB documents into their context.

[עברית](README.he.md)


Why

Big projects grow documents faster than any agent can read them: a 200KB plan, a 300KB handoff log, an 80-row decision table, rules spread over CLAUDE.md files. An agent that reads them whole burns its context, still misses the one line that matters, re-opens decisions settled weeks ago, and loses the resume point at every compact. Telling it "don't read big files" lasts until the first compact.

CIS changes the habit:

	without	with
getting oriented at session start	read SESSION-CONTEXT / handoff files (~50KB)	a ≤4KB card injected automatically
"where did we decide X?"	open and skim the decision log	cis ctx "X" ⇒ 5 refs + the exact lines, ≤3KB
reading a 300KB doc "just in case"	allowed	blocked — read the range ctx points to
editing billing.ts	no idea a decision governs it	≤3 lines: the decision / rule / memory that names it
keeping it current	someone remembers to	git hooks rebuild the index in the background

Zero dependencies — Node 18+, Windows / macOS / Linux.

Install

One repository, three ways in — pick by your setup (check your version with `claude --version`):

**Claude Code 2.1.275 or newer** — one command:

/plugin install central-info-station --marketplace rordan-ai/central-info-station