← Business Idea Generator · Prompt Deck · Pictures

How does a platform turn "I know fishing charters" into an operating business — and get smarter each time?

Build Contract​

The deliverable, not part of the framework. The Tight Five sections below justify what's in this table. This is what engineering builds from and what commissioning reads.

Principles​

What truths constrain the design?

The Job​

The hardest thing: The 95% commodity that has nothing to do with their expertise. Not the domain — the everything-else.

Why Now​

• AI displacement accelerating — millions of people will need to rethink how they make a living in the next 2-3 years
• Sprout Generator exists (built, migration applied) — infrastructure layer waiting to be called
• Nine algorithms designed — intelligence layer ready for implementation
• venture_business_ideas table already in schema — the entry point is architected
• Business development framework documented — the thinking has been done, the code hasn't

Design Constraints​

Refusal Spec​

Performance​

How do we know it's working?

Priority Score​

PRIORITY = Pain x Demand x Edge x Trend x Conversion

Quality Targets​

Failure Budget​

Eval Strategy​

Kill signal: If the first 5 ventures created by the generator produce zero WorkChart receipts within 30 days of launch, the generator is producing shells not businesses. Stop and investigate whether the problem is quality (bad intelligence) or completeness (missing operational setup).

Platform​

What do we control?

Current State​

Input Universe​

Human Fallback​

Cost, Latency, Quality​

Build Ratio​

~70% composition (schema + Sprout + shared libs + algorithms), ~30% new code (orchestration pipeline, AI research prompts, template generation logic)

A2A Infrastructure (Honest Assessment)​

What the codebase calls "A2A" is internal venture-to-venture routing, not the Google A2A protocol. The Business Idea Generator inherits this from Sprout. The interfaces are HTTP, the discovery path is standard (.well-known/agent-card.json), the agent card concept is right. What's missing is schema conformance, task lifecycle, and cryptographic trust.

Position: Build ventures that work internally now, shaped so they're A2A-conformant when they go external. Don't implement the protocol today — there's no external network to talk to. But don't build in ways that make conformance a rewrite.

Parked until: external agents exist that ventures need to communicate with.

Protocols​

How do we coordinate?

Build Order​

Commissioning​

Agent-Facing Spec​

Commands:

pnpm test:unit       # Algorithm unit tests
pnpm test:e2e        # End-to-end: idea input → venture created
pnpm eval:research   # Eval-runner against domain research quality
pnpm eval:valueprop  # Eval-runner against value proposition quality

Boundaries:

Test Contract:

Players​

Who creates harmony?

Demand-Side Jobs​

Job 1: Craftsperson Starts a Business​

Situation: A skilled person — displaced by AI, retiring from corporate, or simply ready to go independent — decides to turn their craft into a livelihood.

Features that serve this job: #1, #2, #3, #4, #5, #6, #7, #8

Job 2: Platform Engineer Creates Venture for Client​

Situation: A builder wants to help someone launch a business on the mycelium but infrastructure wiring takes weeks of commodity work.

Features that serve this job: #6, #7, #9, #10

Example Pairs​

Role Definitions​

Venture as Universal Work Unit​

A venture isn't just "a company." It's the universal unit of scoped work at any scale:

The generator adapts to scale. parentVentureId in the schema enables nesting.

Relationship to Other PRDs​

Context​

• PRD Index — Business Idea Generator
• Prompt Deck — 5-card pitch
• Pictures — Pre-flight maps

Questions​

Which requirement in the business idea generator spec is most likely to change after the first user test — and is the architecture resilient to that change?

• Which acceptance criteria are genuinely verifiable at commissioning time vs which require subjective judgment?
• If the spec were handed to a new engineer with no context, which part would produce the most questions?
• Which requirement in this spec is load-bearing — the one where being wrong propagates failure to all others?