The Instinct Transfer
The document is a workshop‑style manual authored by Mario Castelli that teaches creative strategists how to move their 15‑year copywriting instincts into an AI using Claude Code. It explains that instincts are hidden in the edits a copywriter makes, and that by capturing those edits as data the AI can run the copywriter’s taste as a checklist before the copywriter ever sees a draft. The system consists of three text files – a Taste file (examples of best work), a Rules file (hard rules and tendencies) and a Pairs file (before/after edit pairs) – plus a standing "Judge" instruction that forces the AI to grade drafts against the Rules file first. The guide provides a step‑by‑step protocol: set up a Claude Code workspace, seed the Taste file, mine before/after pairs to extract failures, principles and tendencies, tier tendencies into hard rules only after recurrence or explicit approval, install the Judge, and adopt daily habits ("save that", "that worked", end‑of‑session sweep) to continuously capture new instincts. It also details safeguards against drift (pure taste file, regular audits, scheduled red‑pen sessions) and a second‑layer of advanced lessons learned from applying the system to emails, short‑form video scripts, and ad hooks (job‑question priority, using picks as training, grading ideas naked, avoiding formula traps, cold judging, tagging rules by creator vs format). The guide outlines eligibility criteria for any deliverable, a 30‑day rollout timeline, and includes appendix prompts for each automation step. The overall claim is that after a few weeks the AI can produce drafts that already satisfy the copywriter’s taste, freeing the copywriter to focus on higher‑level creative decisions.
Topics
Instinct Transfer Overview: Teaching AI Your Taste via Edits
Introduces the core concept that a copywriter's instincts are embedded in their edits and can be captured to teach an AI to run the copywriter's taste as a checklist.
Three‑File Architecture + Judge: Taste, Rules, and Pairs
Details the system's four components – the Taste file (best examples), the Rules file (hard rules and tendencies), the Pairs file (before/after edit pairs), and the Judge instruction that enforces rule checking before drafts are shown.
Step‑by‑Step Protocol: Setup, Seeding, Mining, Rule Tiering, Judge Installation
Provides a chronological workflow: setting up Claude Code, populating the Taste file, mining edit pairs to extract failures and principles, promoting tendencies to hard rules, installing the Judge, and establishing daily capture habits.
Capturing Instincts from Edits: Before/After Pair Mining
Explains how to feed the AI before/after versions of copy, prompting it to name the failure, the underlying principle, classify context, and propose a one‑line tendency for the Rules file.
Rule Hierarchy: Tendencies vs Hard Rules and Promotion Process
Describes the two‑tier rule system where a single edit creates a tendency, and a hard rule is only created after recurrence (3+ instances) or explicit user approval, with blind‑testing as a validation tool.
Operational Habits: "Save That", "That Worked" and End‑of‑Session Sweep
Introduces daily trigger phrases and a 60‑second end‑of‑session review to capture corrections and wins in real time, ensuring continuous feeding of fresh taste data.
Drift Prevention Strategies: Pure Taste File, Regular Audits, Red‑Pen Schedule
Outlines four defenses against the system degrading over time: keeping the taste file pure, the user being the sole rule source, scheduled red‑pen sessions even when drafts look fine, and quarterly rule audits.
Second‑Layer Lessons Learned: Job Question, Picks as Training, Naked Idea Grading, Formula Trap, Cold Judging, Rule Tagging
Summarizes six advanced insights discovered while applying the protocol to emails, video scripts, and ad hooks, covering priority of the job question, using picks for fast training, separating idea evaluation from craft, avoiding over‑reliance on a single formula, judging drafts cold, and tagging rules as creator‑specific or format‑specific.