Medium: Writing PRDs for AI products — a framework for probabilistic systems
What the article is about
Nima Torabi, a product and strategy leader with experience at Samsung and Rogers Media, published a detailed guide on Medium in April 2026 addressing a practical gap: traditional PRDs assume deterministic code, stable data, and a narrow set of stakeholders. AI systems operate under different conditions — they produce probabilistic outputs, require ongoing data governance, and pull in Legal, Trust and Safety, and Data Science as active contributors to the specification process. The article provides a seven-part framework for writing PRDs that match how AI products actually behave.
Context and structure
The guide is structured around seven sections: a conceptual shift from traditional to AI-aware requirements; a strategic definition phase covering problem-solution fit and golden use cases; a data strategy section addressing provenance, freshness requirements, and permission inheritance; a model behavior section covering latency budgets, evaluation strategies, and error handling; a UX and interaction design section on communicating AI uncertainty and managing latency in the interface; a safety and risk section covering refusal specifications, red teaming, and human-in-the-loop protocols; and a collaboration and lifecycle section describing eval-first development and post-launch monitoring.
Key takeaways
The article’s most practical contribution is the shift from treating the PRD as a handoff document to treating it as an operational contract that evolves through the product’s life. Torabi identifies several requirements that traditional templates omit entirely. A Freshness Requirement Table should document acceptable data age for every source in the system, because without it, engineering will either over-provision real-time syncing or ship a model that reasons from stale inputs. System personas — descriptions of how the AI should behave in tone, refusal patterns, and uncertainty expression — should appear as a primary specification, not a designer afterthought. Acceptance criteria should be machine-verifiable and probabilistic rather than qualitative terms like “intuitive” or “fast.”
Safety surfaces get their own section, not a paragraph in the appendix. Torabi argues that refusal specifications, hallucination tolerance thresholds, and escalation paths to human review are product decisions, not engineering decisions, and belong in the PRD alongside features.
The section on token economics addresses a common gap in discovery: understanding cost at query volume matters before committing to an architecture, and product requirements should define cost targets alongside performance targets.
Who it is useful for
Product managers who have shipped traditional software and are now responsible for AI-powered features that have been poorly specified on earlier attempts. Also useful for cross-functional leads in Legal and Trust and Safety who want to understand how AI requirements differ from software requirements and where their input shapes product decisions. The seven-part structure makes it practical to adopt partially — a team can apply the data strategy and model behavior sections without rewriting its entire PRD template.