§ reference · skill

discover

Interviews you one question at a time to capture feature intent (Goals · Non-Goals · Functional Requirements · Acceptance Criteria · Decisions) into a Feature Requirements Document the research skill consumes.

arguments [free-text feature description | existing artifact path]

§ 01 · purpose

Purpose

The canonical entry point when the feature idea is still fuzzy. discover pins down intent through a one-question-at-a-time interview so subsequent agents do not chase the wrong target. The FRD's Decisions block is consumed by research and propagates through Developer Context into design.

§ 02 · when to use

When to use it

  • The idea is fuzzy and the team wants it stress-tested before any codebase probe.
  • You have a half-written ticket or rough spec to refine into a structured FRD.
  • Skip when you already have a clear spec or ticket — go straight to research.

§ 03 · inputs

Inputs

Name Required Source
$ARGUMENTS yes Free-text feature description OR path to an existing FRD/ticket/doc to refine
$ARGUMENTS
A path triggers refinement mode — file is read FULLY as baseline context.

§ 04 · outputs

Outputs

Artifact Path Format
Feature Requirements Document .rpiv/artifacts/discover/ markdown (research-compatible)

§ 05 · key steps

Key steps

  1. Foundational intent question first — no agents, no file:line Why: Intent shapes the probe scope. Probing the codebase before stated intent risks framing the FRD around what exists rather than what the developer is trying to solve.
  2. Lightweight codebase probe shaped by stated intent Why: Parallel locator agents run only after the intent answer narrows the slice. Keeps probe cost proportional to the feature size, not the codebase size.
  3. Build the decision tree lazily — root + immediate children Why: Expanding one layer at a time avoids speculative questions that depend on answers not yet given.
  4. Batch-confirm evidence-based pre-resolutions Why: When the probe surfaces a likely answer, the agent proposes it for confirmation rather than silently recording it. Surfaces disagreement before it propagates downstream.
  5. Interview loop — tiered questions, re-queue cross-cutting answers Why: Each answer can spawn follow-ups; cross-cutting answers re-queue affected branches so the tree stays internally consistent.
  6. Synthesize answers into FRD sections; write a fresh artifact Why: Each invocation always writes a NEW timestamp-distinct artifact — never appends — so a prior FRD is never silently mutated mid-iteration.

§ 06 · related skills

downstream researchexplore