startup-team-marketplace

Startup Team Skill

An agent skill that simulates an adversarial leadership team to stress-test your startup and product ideas before you commit to building them. Works with any agent that supports skills.sh — including Claude Code, Cursor, Codex, Cline, Windsurf, Gemini, GitHub Copilot, and more.

Run /startup <your idea> and a team of domain-expert personas — from Head of Product to Principal SRE to Head of Security — will debate, challenge, and refine your idea across multiple rounds. 15 personas are built in, and the skill dynamically creates new ones when your idea touches a domain they don't cover. The result is a structured decision document and, if the team recommends proceeding, a complete product requirements document ready for engineers or agent teams to start building from.

Installation

Prerequisites

• Any skills.sh-compatible agent installed and configured
• The OpenProse plugin installed (execution engine for the adversarial loop)

Install

npx skills add bryanmig/startup-team-skill

That's it. The /startup command is now available.

Verify Installation

Open Claude Code and run:

/startup A simple test idea

You should see the team selection phase begin, followed by initial positions and adversarial rounds.

Usage

Basic

/startup <your idea>

With Custom Round Limit

/startup <your idea> --max-iterations 3

The --max-iterations flag controls the maximum number of adversarial rounds (default: 5). Lower values produce faster but less thorough evaluations.

Examples

/startup AI-powered code review tool that learns team patterns
/startup A marketplace connecting freelance CFOs with Series A startups
/startup Build a boring calendar replacement that actually saves time
/startup Developer tool that auto-generates API documentation from code --max-iterations 3

Output

Each run produces a decision document at docs/decisions/YYYY-MM-DD-<slug>.md:

• Idea Summary — the idea refined through discussion
• Participants — who weighed in and their final stances
• Recommendation — PROCEED / PROCEED WITH CONDITIONS / DO NOT PROCEED / PIVOT TO...
• Supporting Arguments — strongest points in favor, attributed to specific personas
• Dissenting Views — who disagreed and why, preserved faithfully
• Risks & Mitigations — failure modes with concrete mitigations
• Vetoes — any veto holders who blocked, and resolution status
• Next Steps — concrete actions if proceeding

If the recommendation is PROCEED or PROCEED WITH CONDITIONS, a product requirements document is also generated at docs/prd/YYYY-MM-DD-<slug>-prd.md:

• Problem Statement — who, what pain, why now
• Target User — specific role, context, workflow
• Success Metrics — with concrete targets and measurement methods
• Functional Requirements — numbered, prioritized (P0/P1/P2), with acceptance criteria
• Non-Functional Requirements — security, reliability, quality, performance, observability
• Architecture Guidance — chosen approach, rejected alternatives, open questions
• Out of Scope — what was explicitly excluded and why
• Implementation Phases — phased delivery with decision gates and veto holder sign-offs
• Traceability Matrix — every requirement mapped to the persona argument that established it

How It Works

/startup AI-powered code review tool that learns team patterns

Phase 0 — Team Selection. A selector agent analyzes your idea and picks 5-8 relevant personas from the built-in 15 plus any previously-created project-local personas. Veto holders are always included when their domain applies. If the idea touches a domain no existing persona covers, the selector flags the gap, asks you to confirm, and a new full-depth persona is created and saved to .claude/personas/ in your project for future reuse.

Phase 1 — Initial Positions. All selected personas respond in parallel with their domain-specific assessment: opportunities, risks, conditions for support, and an initial stance.

Phase 2 — Adversarial Rounds. Personas challenge each other directly. They quote specific arguments, explain why they're wrong, defend when challenged, and update their stances. A facilitator evaluates convergence after each round. Rounds continue until consensus or until disagreements are clearly irreconcilable (max 5 rounds by default).

Phase 3 — Synthesis. The decision owner makes the final call using a Truth Hierarchy to resolve value conflicts:

Reality > Safety > Correctness > Long-Term > UX > Market > Speed > Narrative

Phase 4 — Decision Document. A decision document is written to docs/decisions/YYYY-MM-DD-<slug>.md with the full recommendation, supporting arguments, dissenting views, risks, vetoes, and next steps.