Model Designer

Key activities & domains

• Shape and refine content frameworks — define the principles that govern how MAIBA should behave in conversation (e.g. "answer first, explain second," "one action per turn") for your focus areas, in partnership with core CD, who approves the north-star bar.
• Translate frameworks into model instructions — convert principles into specific, testable system prompts (turning "be actionable" into "always include a specific number and a deeplink").
• Hillclimb on language-driven behavior — iteratively improve any behavior controlled by language rather than code: tone, answer structure, length, follow-up prompts, ambiguity and "can't help" handling.
• Interpret eval results into model-design priorities — read judge scores and hillclimb readouts and decide what to tune next; this is the feedback loop between measurement and design.
• Align partners on the quality bar — get PM, eng, PD, and leadership on the same page about "what good looks like" before the model changes, heading off the "I'll know it when I see it" problem.
• Dogfood and triage — exercise the model, surface language-level failures, and partner with eng to flag "big rock" failures before they reach experiment audiences.
• Operate within the sprint loop — per 4-week sprint: shape/update the focus framework (week 0), write and refine prompt rules (weeks 1–2), and use judge scores to iterate (weeks 3–4).
• Stay consulted on adjacent surfaces — partner (not dictate) on in-response design like widgets, where model handling and visual design intersect; hand Python-explicit behavior (routing, tool calling) to eng.

Required Skillset

• Strong content / UX writing craft and user empathy; reasons about a conversation as an end-to-end experience, not isolated strings.
• Hands-on system prompt / LLM behavior iteration — has written and refined prompts and can articulate why a phrasing change shifts model behavior.
• Strategic judgment — can hold and refine the "what good looks like" question, not only execute against a given spec; able to operate in ambiguity and set direction for a focus area.
• Comfort with shared ownership — works confidently within a defined sign-off gate (core CD approves the north-star bar) while exercising latitude inside it.
• Analytical iteration habit — reads judge / eval scores and uses them to prioritize the next change.
• Partner alignment — can get PM, eng, and design to a shared quality bar before building.
• Fast, high-volume execution under a 4-week sprint cadence.

Nice-to-have

• Technical writing experience;
• Ads / performance-marketing domain familiarity;
• Prior model design or AI conversation design experience;
• Information architecture background.