About this collection

About this collection

Idea Validation Kit — Claude Code / Cowork

1 of 3 in this book · 1 min read

A research-first, evidence-audited multi-agent workflow to screen business ideas before building products. Language: Vietnamese for reports; English identifiers for machine-readable schemas.

Quick start (Claude Code)

  1. Extract the archive, open this folder as a Claude Code project, and read CLAUDE.md.
  2. Copy templates/idea-input.md to outputs/<slug>/idea-input.md and fill what you know. Partial inputs are allowed; missing items become assumptions.
  3. In Claude Code, run /validate-idea outputs/<slug>/idea-input.md (project command under .claude/commands/). Alternatively ask Claude to follow workflows/validate-idea.md.
  4. Review decision-report.md, evidence-ledger.csv, scorecard.csv, red-team.md, and experiments.md.
  5. Do real-world experiments before treating an idea as validated.

Requirements

  • Claude Code with web research tools/search access for externally grounded research. If unavailable, output UNVERIFIED and a research backlog; never fabricate citations.
  • Parallel subagent spawning depends on Claude Code features and permissions. Agents can run sequentially with the same contracts.
  • Claude Cowork: use CLAUDE.md, workflows/validate-idea.md, agent definitions and templates as instructions; don't assume Claude Code slash commands/subagents are natively available.
  • This kit does not call paid APIs, collect user data, or autonomously interview customers.

Structure

  • .claude/agents/: ready-to-use Claude Code subagent definitions.
  • .claude/commands/: /validate-idea project command.
  • workflows/: orchestration sequence, synthesis rules, stop conditions.
  • rubrics/: 12 criteria, weights, evidence quality, scoring/decision principles.
  • templates/: intake, evidence, output, experiments.
  • examples/: fictional sample idea and sample decision excerpt (NOT actual validation).
  • scripts/score.py: deterministic scoring from an evidence-backed scorecard CSV.

Operating principles

Separate facts, inference and hypotheses. Each claim gets dated source/evidence ID; a missing source means unknown, not false or true. No ungrounded TAM, conversion rate, pricing or competitor claims. Never claim success from desk research alone. Prefer small reversible tests over building features.