Discovery

Overview

Discovery is a systematic interview process that clarifies requirements before building. Instead of asking open-ended questions and hoping for complete answers, it proposes answers based on available context and asks the user to confirm or correct. One question at a time, with a proposed answer — faster and more precise than open-ended questioning.

Interview Modes

• Quick Mode (5-10 questions) — for L1-2 tasks where scope is mostly clear. Covers goal, constraints, and done criteria.
• Standard Mode (15-25 questions) — for L2-3 features affecting multiple files. Adds users, edge cases, dependencies, and alternatives.
• Deep Mode (25-50 questions) — for L3-4 major features with business impact. Adds business model, scale, security, integration, and migration topics.

Codebase-First Rule

Before asking any question, check if the answer already exists in code or docs. Sources checked in order: projectbrief, productContext, techContext, systemPatterns, package.json, README, existing code. If you can look it up, state what you found and ask the user to confirm — don't ask from scratch.

Dependency Tracking

Answers can invalidate earlier answers. Discovery tracks dependencies between questions and revisits affected answers explicitly when new information changes the picture. This prevents building on assumptions that are no longer valid.

Pre-Checks

Before proposing port numbers, subdomains, or database names, verify against existing allocations. Port conflicts discovered late in deploy are expensive to unwind. For comparative research tasks, re-read ALL raw materials the user provided at inception — do not rely on session memory.

When It's Used

The planner agent loads Discovery during /dr-prd or before /dr-init for tasks where requirements are ambiguous. The output is a structured requirements summary with goal, scope, requirements table, constraints, edge cases, and done criteria.