The rubric is the product
Annotation taught me that model quality lives or dies in the labeling rubric. Vague guidelines produce vague data: one annotator's "helpful" is another's "harmful." The fix is always the same — concrete examples, named edge cases, and a decision log for the weird ones.
I carry that into red teaming. An eval without crisp verifiers is just vibes; every must and must-not gets a deterministic check where one is possible.
Volume builds instinct
After thousands of TTS and chat judgments across multilingual datasets, you develop a feel for where models wobble: long contexts, conflicting instructions, polite-sounding traps. That instinct is exactly what makes a good adversarial task — I plant attacks where I know the footing is loose.
Both sides of the loop
Annotators feed models; red teamers audit them. Doing both at once — RWS by day, Alice red teaming alongside — means every labeling session sharpens my attacks and every attack sharpens my labels. If you do one, learn the other.