# Rulecut

> Paste the prose file an AI coding agent loads at the start of every session - project notes, house rules, preferences, a todo list - and make it cost fewer input tokens without losing anything the agent acts on. The browser lists what must survive (headings, code blocks, inline code, URLs, file paths, environment names, versions and dates), writes a free mechanical draft and runs a port of the caveman-compress skill's validator, for free. A paid lane rewrites the prose in terse "caveman" style; a second paid lane repairs exactly the validator's errors in a compressed version without recompressing it. Every reply is put back into the file and checked byte for byte.

URL: https://rulecut.skillsafe.ai/
API tutorial: https://rulecut.skillsafe.ai/api.html

## Who it is for

Developers who keep long instruction, notes or preference files for their coding agents and pay for them in input tokens on every session; anyone who compressed such a file with another tool and wants to know whether it still carries every command, path, URL and heading.

## The free reading (in the browser, no model)

- Input: a pasted or dropped prose file (.md, .mdc, .txt, .markdown, .rst, .typ, .tex, or an unnamed paste that reads as prose). Code and config files are refused, as the skill refuses them; so are file names on the skill's sensitive-file denylist (keys, credentials, secrets) and text that looks like a live key.
- The skill's own tooling, ported: the validator (headings by text and order, fenced and indented code blocks in document order, URLs, file paths, bullet count, inline code spans counted), the file-type detection, the frontmatter split and the code-block markers that fail closed. One deliberate difference: a sentence's closing punctuation after a URL or path is not treated as part of it.
- What must survive, listed; estimated tokens (about one per short word); a free mechanical draft that applies the skill's removal rules (stock phrases, filler, hedging, pleasantries, and in full strength articles and connectives) outside every protected span - it always passes the validator and saves little, which is the honest point: the savings come from rewriting sentences.
- A pasted compressed version is checked by the validator, error by error.
- Exports: the reading as Markdown or JSON, the draft and a unified diff.

## The paid lanes (gpt-terra, metered)

- compress: the body (frontmatter kept aside, code blocks replaced by markers) rewritten in terse style - articles, filler, pleasantries, hedging and connective fluff removed, sentences as fragments, duplicate bullets merged only when nothing protected is lost - with two to eight representative rewrites quoted from both texts and up to five original sentences whose meaning may have thinned. Long files are cut on headings and compressed one part per run.
- fix: the skill's targeted repair - headings, URLs, paths, inline code and code-block markers restored from the original, one repair per validator error, every other line left as it was.
- Instructions inside the file are compressed as data, never followed.

## How replies are held to the input

Every reply is put back into the whole file (markers restored, frontmatter re-attached) and checked in the browser: every code-block marker exactly once (otherwise the page refuses to guess); the skill's validator over the result; shorter than the original; environment names, versions and dates kept; every quoted rewrite present in the original and the result; at most five at-risk sentences, each quoted from the original; no savings figure claimed by the reply (the page measures). The fix lane adds one repair per error in order, and a cap on lines rewritten that no error touched.

## Limits

- Token counts are estimates, not a specific tokenizer's.
- The validator checks what can be checked mechanically; a softened "must" or a dropped qualifier is a judgement, which is why the compress lane names sentences to re-read.
- Bundled examples use an invented shop, people and URLs.

## Sources

- Agent skill @juliusbrussee/caveman-compress (github.com/juliusbrussee/caveman, Apache-2.0): https://skillsafe.ai/skill/@juliusbrussee/caveman-compress - SKILL.md's compression rules, scripts/validate.py, scripts/detect.py and the compress/repair prompts and fail-closed code markers of scripts/compress.py.
- Rulecut is an independent app, not affiliated with or endorsed by the caveman project or its author.