Your agent rereads this file every session. Make it cost less.
Paste the notes, house rules or preferences file your coding agent loads at the start of every session. Your browser lists everything that must survive word for word - headings, code blocks, inline code, URLs, file paths, environment names, versions and dates - writes a free mechanical draft, and runs the caveman-compress skill's own validator. Nothing is uploaded until you run a lane. Then Compress rewrites the prose in terse caveman style, and Fix repairs exactly what the validator flags in a compressed version, without rewriting the rest. Every reply is put back into the file and checked byte for byte.
The shop, the people and the URLs are invented. Each example comes with a saved model reply, so you can see the whole page without signing in or spending a credit.
What this does, and what it does not
An AI coding agent reads its instruction files - project notes, house rules, preferences, todo lists - at the start of every session, so every sentence in them is paid for in input tokens again and again. Most of that prose can lose its articles, filler, hedging and politeness without losing a single rule. What it must never lose is the part the agent acts on: the commands, the paths, the URLs, the exact names in backticks, the headings other files link to, the versions and dates. The skill behind this page compresses the prose and then validates the result against the original before it is allowed to overwrite anything.
This page runs that validator in your browser - a port of the skill's own, with one deliberate difference: a full stop after a URL is not part of the URL. It keeps the frontmatter aside, replaces every code block by a marker before sending (so code is never sent to be rewritten, and comes back byte for byte or the page refuses), writes a free mechanical draft, and estimates the tokens before and after. The Compress lane rewrites the prose in terse "caveman" style and names its representative rewrites and any sentence that may have thinned; the Fix lane does the skill's targeted repair: only the errors the validator lists, nothing recompressed.
Files the skill would never send - names on its denylist of keys, credentials and secrets, or text that looks like a live key - are refused before anything leaves the browser, and code or config files are refused too: the skill only compresses prose. Instructions inside your file are compressed, never followed. Token counts are estimates (about one token per short word); the savings figures are measured by the page, never claimed by the model. Derived from the agent skill @juliusbrussee/caveman-compress (github.com/juliusbrussee/caveman, Apache-2.0: the compression rules, the repair step, the validator and the file detection). Rulecut is an independent app, not affiliated with or endorsed by the caveman project or its author. Sources are listed in llms.txt.