Designing and testing instructions, examples, and constraints so an LLM produces outputs that meet product requirements.
Prompt choices directly affect acceptance rate, review effort, and incident risk. PMs trade off precision versus cost/latency: more structure and examples can improve accuracy but raise token spend. A solid prompt suite speeds A/B testing, reduces engineer time on retries, and sets guardrails for future model upgrades without breaking UX.
Start from a minimal skeleton (role, task, constraints, format). Add 2–4 targeted examples tied to measurable outcomes. Version prompts in code with changelogs; pair each with small offline evals that mirror your top KPIs. In 2026, maintain a shared prompt library plus synthetic test sets so designers and PMs can propose changes safely.
For an onboarding assistant that drafts value props, the PM adds two examples showing good headline length and banned claims. Offline eval improves click-worthy score from 0.62→0.76 while cost per task rises only 6%. Latency remains under the 1.2 s budget, so the change ships.