Notebook Entry 017
"Review against" ≠ "read the file" — voice drift in AI-drafted content
April 13, 2026
Why did a drafted comment sound nothing like how Amanda actually talks?
What happened
Claude drafted a LinkedIn comment without reading voice-and-style.md, anti-ai-writing-style.md, or content-identity.md. The comment read like a consulting textbook — generic analysis, no personal voice, multiple anti-AI structural tells ("The fix isn't better messaging. It's better sequencing" = mic-drop formula). When Amanda flagged it, Claude admitted it hadn't read the files. This directly undermines Amanda's "human-first AI implementation" positioning.
Root cause
CORE-00 said "review against REF-anti-ai-writing-style.md" — which Claude interpreted as "keep the general idea in mind" rather than "read this file right now." The word "review" is ambiguous. Additionally, no mechanism forced re-reading the content files at the point of drafting — even if they were read during boot, they'd be thousands of tokens back in context by the time a LinkedIn comment is being written mid-session. Rules that say "do X" without enforcement mechanisms are wishes, not rules.
What changed
CORE-00 Rule 6 rewritten from "review against" to "MUST read these three files using Desktop Commander — not from memory." Added mandatory pre-flight checklist (Would Amanda say this? Anything specific to her? No-fly-list scan? Structural tell scan? Lane-two layer for comments?). Same rule added to system constitution for project-level searchability. The word "review" is banned from file-read instructions — replaced with "read using Desktop Commander."
The lesson
In multi-agent AI systems, "soft" instructions ("review," "consider," "keep in mind") are systematically ignored under time pressure or context window limitations. AI assistants take the path of least resistance — generating from training data rather than reading reference files — unless the instruction is explicit, mandatory, and verifiable. For any organization using AI to draft brand-sensitive content: your style guide must be read at the point of generation, not just loaded at session start. And "review" must be replaced with "read" in every instruction.
One implementation insight every other Tuesday.
What actually works when you ship AI in real organizations. Nothing you could get from a press release.