"Your AI keeps losing its voice" — @ship30for30 (Cole Schafer)

Why this is in the vault

Identifies the root cause of AI voice drift — context pollution from mixing format types in a single thread — and names a simple structural fix directly applicable to RDCO's AI-assisted content workflow.

The core argument

The newsletter is a weekly digest of three social posts. The AI voice piece (item #1, attributed to Cole) makes one crisp claim: voice inconsistency in AI-generated writing is caused by context pollution — specifically, loading newsletters, social posts, and marketing copy into the same conversation thread forces the model to reconcile competing register rules, and the output averages out into something that sounds like none of them.

The implied fix (stated as "Cole's simple fix") is format isolation: keep each content type in its own thread or context window so the model holds exactly one set of voice rules at a time. The email doesn't elaborate further — the full breakdown is behind the linked social post.

Secondary items in the digest: Pixar's 22 storytelling rules (story spine formula) and Cole's 7 copy sentences that "make customers sell themselves." Neither is developed in the email body; both link out.

The PS is a hard sell for the "Personal Agent Bootcamp" ($800, enrollment closing June 22), co-run by Dickie, Cole, and Mitch. Pitch: build a personal AI agent, productize it at $3k–$5k/mo.

Mapping against Ray Data Co

The format-isolation principle maps directly to RDCO's content stack. The Sanity Check newsletter, short-form social (X threads), and any marketing copy for Squarely or MAC are three distinct registers. If Claude is being used to draft across all three in the same session context, this is exactly the failure mode described — the model starts blending tones.

Practical implication for RDCO: use dedicated subagent dispatches (or separate Claude sessions) per format type when producing content. The COO agent already runs skill-scoped subagents; extending that discipline to content drafting is the direct application.

The bootcamp pitch is tangential — RDCO is building its own agent infrastructure, not buying into a $800 cohort course.

Related

• [[~/rdco-vault/02-sops/2026-06-09-claude-md-prompt-precedence-full.md]]
• [[~/rdco-vault/CAPABILITIES.md]]