Fontai CRUX

Get to the crux of long documents.

You got a 40-page analysis and have to decide. CRUX reads it for you and answers on one screen: a verdict, what matters, what is missing, what to ask the author. It also checks whether the analysis can be trusted.

What you get

One screen instead of forty pages.

Built for people who get long, often AI-written analyses and have to make a decision based on them.

01

Verdict first

Approve, Revise, Return or Missing info, with one sentence why. Many readers stop here.

02

The summary

Main message, key points, risks and gaps, and questions worth asking the author. Each point links to its page or section.

03

A quality score

Seven criteria on a 1–5 scale, each backed by evidence. Key claims are checked on the web, and cited sources are checked to see whether they exist.

04

The brief behind it

What was the author probably asked to do, and why? It helps when you don't know where the document came from.

How it works

Say “crux” and hand it the document.

Markdown, Word, PDF or pasted text. CRUX replies in your language, whatever the document's language.

1

Give it the document

crux report.pdf, or attach the file and write “crux”.

2

Answer four questions

What to do, how deep, the author's brief (if you have it), and what you need to decide.

3

It reads and checks

In Claude Code, separate helpers do the research and return only short findings, so your session stays light.

4

You decide

The answer stays in the chat. Ask if you want it saved or exported to PDF.

How deep

You always know how long it takes.

Pick the depth before it starts. Long documents (50+ pages) take about half as long again.

Quick~1–3 min

Only the document itself: consistency, gaps, typical AI defects. No web.

Standard~5–10 min · recommended

Plus a web check of the claims the conclusions rest on, and of the sources it cites.

Deep~15–30 min

Plus independent research by parallel helpers, and every source checked.

Example answer

What it looks like.

A short test analysis recommending a switch of CI platform. It had a made-up source and wrong prices planted in it, and CRUX found them all. Shortened.

VERDICT: Return. The savings are roughly doubled and the performance argument rests on a source that cannot be found. Brief: inferred (“recommend whether to move from GitLab CI to GitHub Actions”) · Your decision: approve the migration, yes/no SUMMARY Main message: move to GitHub Actions in Q1 for ~110,000 CZK/year in savings and 37 % faster builds (§1, §2, §3). RECOMMENDATION — for your decision • Don't approve yet. Ask for the GitLab invoice and real CI-minute usage. • From current price lists the saving is ~53,000 CZK/year, not 110,000. RISKS AND GAPS ⚠ Savings contradict each other: §2 says 110,000, §5 says 80,000 ⚠ GitHub Free has 2,000 Actions minutes, not 10,000 (§2) [3] ⚠ The cited Gartner report cannot be found — likely made up (§3) ⚠ The “Risks” section is filler only (§4) QUESTIONS FOR THE AUTHOR 1. Where do 120,000 CZK and 10,000 minutes come from? 2. Can you share the Gartner report? SCORING (Standard · brief inferred) Total 1.8/5 → Return Accuracy 1 · Evidence 1 · Specificity 2 · Consistency 3 Usefulness 1 · Depth 1 · Alignment 2 (indicative) [1] about.gitlab.com/pricing [3] github.com/pricing
Install

Free. Two minutes.

CRUX is a skill for Claude. It works in Claude Code and in the Claude app on the web, desktop and phone.

1

One click, if you use Claude Desktop or VS Code:

Claude Desktop opens with the install request filled in. Just press Enter.

or in a terminal, one line:

claude plugin install crux --marketplace fontai/crux

or paste this into any Claude Code session and let Claude do it:

Read https://github.com/fontai/crux/blob/main/README.md and install the crux skill.
2

Start a new Claude Code session and try it: crux path/to/document.pdf

What CRUX promises

Strict, and never makes things up.

No made-up sources

It cites only pages it actually opened. Everything else is marked unverified.

Strict scoring

AI tends to go easy on AI-written text. CRUX needs the same evidence for praise as for criticism.

Your data stays yours

Web searches use generic terms about the topic, never names or numbers from your document.

Always up to date

It checks once a day and tells you when a new version is out.