# AI-feel instinct turned into a mechanical rule (L0008)

User's reasoning in Exercise 0002 included a taste-based rejection of the Contrast headline pattern: "feels AI-generated." The instinct was correct but not actionable; the user could not specify *what* made the pattern feel AI-shaped.

L0008 converts the instinct into a checklist. Eight tells (em-dashes, triadic abstract lists, empty intensifiers, "not just X, it's Y", formal transitions, vague unfalsifiable claims, symmetric balanced sentences, lack of proper nouns and numbers); the replacement test ("strip every noun and replace with 'thing'"); three filters humans add (specificity, position, length variation); a 5-step audit.

Note the meta-loop: this workspace already eliminated em-dashes by user rule (NOTES.md), which is also the single highest-signal AI tell. The taste preference and the mechanical rule align. Future content created in this workspace inherits one tell-elimination by default.

For BI / Holistics writing specifically: the lesson emphasises proper nouns (Carousell, Tiki), specific numbers (500M-row table, 2,400 dashboards), and falsifiable claims as the fixes that matter most. The user works in presales; their natural source of specificity is the prospect names and metrics in their actual deals. The audit gives them permission to write what they would say on a call, not what a marketing template tells them to write.

Pattern for future critiques: when scoring writing, run the 8-point audit alongside the L0006 design checklist. Two-layer review (visual + voice) catches both classes of weakness.