Shuten

Shuten

Two plugins that cut an AI agent's excess text: **Shut** — what it writes to you in the chat, **CCShut** — what it leaves as comments in your code. Both load at session start, so they apply from the first message rather than from the moment the agent decides they are relevant.

English version: the question, the warning and the final answer are plain English. Ukrainian version — [Shutuk](https://github.com/Niedvin/Shutuk).

They install into Claude Code and Claude Desktop as plugins, and into Codex, opencode, Gemini CLI and Cursor as skills.

---

What they actually do

Shut — token savings

The problem: the agent spends your tokens on text that changes nothing — reports on its own process, announcements of "now I will read the file", restatements of your request, half-screen summary tables. You pay for it and do nothing with it.

Shut allows exactly four shapes of message and nothing else:

Shape	Register	Rule
Step label	telegraphic	≤ 4 words, no comma or dash inside, most often absent entirely
Question	plain	only when your request can be read two ways and the two would give different results
Warning	plain	only when you must decide or act right now
Answer	plain	≤ 8 lines, no tables, no headings, no list of the steps taken

No running commentary, no announcing a tool call right before making it, no "run it and see how it works". Numbers as digits, relations as symbols (`>`, `≈`, `→`, `∵`).

**It does not touch the code or the result of the work.** The rule covers only what the agent writes to you.

Hooks re-assert the rule after every skill load and check the turn before it closes, so it does not dissolve over a long session.

CCShut — against context rot

The problem: the agent buries a file in comments that explain what the code already says. A few edits later the code has moved on and the comment has not — now it lies. Every later read of that file, by you or by the agent, costs tokens and