Train system design requirement-discovery with AI using vague-problem drills

anirudhology · reddit · 2026-11-14

The author shares a repeatable AI drill for the skill most system design practice skips: uncovering requirements, constraints and invariants — since your design is only as good as what you discover.

The method:

Ask AI for a deliberately vague, incomplete system-design problem (e.g., a notification system over email/SMS/push) that withholds requirements until you ask the right questions.
Probe actively: are notifications time-sensitive? Normal vs. peak volume? Delivery guarantees? What happens on lost or duplicate notifications?
For each question, note what you're uncertain about and what decision the answer informs — e.g., peak write rate decides single DB vs. write partitions.
Sort everything into known facts, assumptions, and unknowns. The author warns that unrecorded assumptions are exactly how hidden constraints sneak into designs.
Submit discovered requirements, remaining assumptions, system guarantees/invariants, proposed design, and residual risks.
Have AI review missed constraints, overly broad questions, and contradictions — but first ask it to restate the full case, since it may invent details mid-conversation. AI is a reviewer, not the final authority.
Optional: repeat with one secretly changed condition; beginners can keep a question cheat list.

Original post →

Train system design requirement-discovery with AI using vague-problem drills

More from coding & agent