The Agentic Email Engine
The document explains Mario Castelli’s “Agentic Email Engine”, a three‑part AI system for producing marketing emails that sound like his own voice while avoiding generic AI output. The system separates the universal writing craft (ENGINE) from the personal voice and material (PERSON) and from a rule‑based quality filter (JUDGE). Core files (core.md, feed.md, proof.md, judge rules, skill.md) store Mario’s worldview, real‑world anecdotes, verified proof, and hard‑fail rules that block AI‑tells such as “you’re not X, you’re Y”, banned words, vague pronouns, and over‑claims. A loop feeds Mario’s voice notes and weekly life updates into the engine, which drafts emails; a second AI judges them against the rules; Mario edits the surviving drafts, and each edit is reverse‑engineered into a new rule, making the engine progressively more like him. The paper walks through six real email examples (announcement, Jason Statham, Loser, etc.), showing raw drafts, Mario’s edits, and the lessons learned, illustrating how the system learns stance, concrete scenes, one‑sentence‑per‑line formatting, CAPS emphasis, and disarming humor. It also describes the broader roadmap: scaling the system to long‑form content, automating market and business feeds, and enabling others to fork the universal engine while keeping their personal voice private. The overall claim is that the moat is not the AI itself but the captured personal voice, material, and rule‑based judge, which together produce high‑scale, high‑quality, non‑generic copy.
Topics
Three‑Bucket Architecture: Engine, Person, Judge
Defines the separation of universal writing craft (ENGINE), personal voice/material (PERSON), and rule‑based quality filter (JUDGE) into distinct markdown files.
Core.md – The Immutable Voice Fingerprint
Stores Mario’s worldview, humor style, recurring characters, and the “coin” CAPS convention that gives each email a unique, repeatable hook.
Feed.md – Live Life & Market Feed as Moat
An append‑only journal of real events, wins, numbers, and market conversation that feeds the engine with uncopyable, up‑to‑date material.
Proof.md – Verified Evidence Rulebook
A strict ledger of real proof (names, numbers) with a hard rule that only verified rows may appear in copy, preventing fabricated claims.
Judge Rules & Hard‑Fails – Anti‑AI Tell Filters
A list of universal disallowed patterns (e.g., “you’re not X, you’re Y”, banned buzzwords, vague pronouns, over‑claims, dashes) that automatically bounce drafts.
Skill.md – The Loop Wiring Diagram
Describes the end‑to‑end workflow: voice note → writer AI (engine+person) → judge AI → Mario edit → rule extraction → next iteration.
Edit‑Driven Rule Learning – The Flywheel
Explains how each of Mario’s manual edits is codified into a new judge rule, making the engine progressively more like his voice and reducing future edits.
Email Case Studies – From Draft to Ship
Six detailed email examples (Announcement, Jason Statham, Loser, etc.) showing raw AI drafts, Mario’s edits, and the specific copy principles learned from each.