discover

Claude Code Discover

A Claude Code plugin that structures product context in your repo before implementation begins. Hypotheses, validation results, and PRDs live alongside your code — so when Claude builds your feature, it has access to rejected alternatives, unvalidated assumptions, and the evidence behind each decision.

Works standalone, or paired with claude-code-workflows for a full discovery-to-implementation cycle:

[claude-code-discover]  →  PRD + Prototypes  →  [claude-code-workflows]
   Discovery phase                                Implementation phase

The Problem

When you ask an AI coding assistant to build a feature, it generates code without knowing what alternatives were ruled out, which assumptions are untested, or what user research shaped the requirements. Discovery artifacts typically live in Notion, Figma, or Slack — invisible to your coding tools. This plugin brings them into the repo where Claude can read them.

What It Does

Vision & Personas        ← who you're building for and why
      ↓
  Opportunities          ← your hypotheses structured with validation plans
      ↓
  Blueprint              ← IA, user flows, content model, brand direction + visual tokens
      ↓
  Hypothesis Files       ← testable assumptions with success/failure criteria
      ↓
  Validation             ← assumption decomposition + HTML prototypes
      ↓
     PRD                 ← each user story traced to evidence

Each recipe is a step in this cycle. Run them in order or jump to where you need:

Recipe	What it does
/discover:recipe-vision	Define product vision, outcomes, and North Star Metric
/discover:recipe-persona	Create personas with JTBD, pains/gains, and behavioral data
/discover:recipe-discover	Structure your hypotheses into Opportunities with validation plans
/discover:recipe-blueprint	Define structural design foundation — IA, user flows, content model, brand direction with visual tokens, AI interaction model
/discover:recipe-refine-visuals	(Optional) Design expert refines auto-derived visual tokens in brand direction
/discover:recipe-validate	Decompose assumptions, design falsifiable tests, generate HTML prototypes
/discover:recipe-reflect	Extract learnings, promote knowledge across the hierarchy
/discover:recipe-define	Generate a PRD from validated hypotheses with confidence scores

What each recipe produces

• Hypothesis file: Markdown with assumption statement, success/failure criteria, confidence scores per risk dimension, time budget, and validation results
• Blueprint artifacts: Information architecture, user flows, content model, brand direction (with auto-derived visual tokens), and AI interaction model — shared structural context that prototypes reference for consistency. Visual tokens (colors, typography, spacing) ensure prototype-to-prototype visual consistency
• Prototype: Single self-contained HTML file (~800-1200 lines) that opens in a browser. Deterministic mock data, all UI states implemented, design context applied from blueprint and project files
• PRD: 200-400 line document with user stories (each with 4 Risks confidence table), EARS-format acceptance criteria, unvalidated assumptions section, and references to hypothesis files

How Validation Works

Each validation produces:

• Assumption breakdown (ranked by risk type and level)
• Test design per assumption (smallest test that could disprove it)
• HTML prototype (for user testing)

Two agents work in separate contexts:

1. hypothesis-verifier decomposes your hypothesis into assumptions, ranks them by risk, and designs the smallest test that could disprove each one — without seeing your expectations
2. prototype-generator reads your design principles, persona, hypothesis files, and blueprint artifacts (when available), then generates a self-contained HTML prototype with deterministic mock data and all UI states

The context separation is deliberate. The verifier designs tests that can fail. The prototype generator builds a product UI without test infrastructure leaking in.

Connecting to Implementation

The PRD that recipe-define produces follows the standard structure that claude-code-workflows expects. Discovery extensions (hypothesis references, 4 Risks confidence per user story, unvalidated assumptions) are additive — they provide context without breaking compatibility. Prototypes generated during validation can be passed to the UI Spec designer as design references.

# Discovery phase (this plugin)
/discover:recipe-define → docs/prd/feature-prd.md