Demon Guide is a legal document drafting tool designed to address the problem of hallucinated citations in legal briefs. Unlike chatbot-based legal assistants, Demon Guide does not generate text through large language models that invent case law. Instead, it uses proprietary models that are not trained on user data. The tool follows a structured, four-pass process that begins with a review of all required papers, then writes, cite-checks, and binds the document. It does not file any brief that contains a fictitious authority. The system is built for attorneys who need source-checked, formatted legal filings that can survive opposing counsel and court scrutiny. The primary workflow involves several distinct stages. First, the user enters structured fields for court, parties, facts, exhibits, and job details. The tool then presents a "Gate" containing every paper the filing actually needs. The "Draft" stage initiates a writing pass, producing a draft without smoothing over missing citations. Two independent "Ward" checkers, who cannot rewrite the brief, evaluate each citation and assign a status of Hold, Stretch, or Kill. The final "Bindery" fetches the current court rules and stamps a formatted PDF and Word document according to those specifications. No human editing occurs during the checking process, and the reference files are wiped after the formatted copy is approved. Demon Guide also provides automatic pleading paper formatting for courts that require it, including numbered line paper for jurisdictions such as California, Arizona, Nevada, and Oregon. The system extracts court-specific rules for paper type, caption, typeface, spacing, and line numbers. It can generate a state box caption, Times New Roman 12 point text, 1.5 line spacing, left alignment, line numbers, left rule, and certificate of service as required by the relevant court rules. The tool is accessible at demonguide.com and is intended for legal professionals who need reliable, court-ready briefs without uploading raw client material to frontier model training datasets.