An AI product-manager agent that turns rough product ideas into structured specs through a 3-phase pipeline: Discovery → Scoping → Spec Writer. It runs a PM-style discovery interview, proposes a RICE-scoped MVP with comparable products, then writes a phased product spec you can hand to a code-gen tool like Emergent, Lovable, Antigravity etc.

Vibe-PM

One conversation replaces weeks of PM process -- from vague idea to phased, prioritized product spec.

Vibe-PM is an AI product-manager agent that turns rough product ideas into structured, developer-ready specs through a 3-phase conversational pipeline. It conducts a PM-style discovery interview, proposes a RICE-scored MVP scope backed by competitive web research, negotiates feature cuts with the founder, and produces a phased Markdown spec you can hand to a developer or feed into an AI code-generation tool.

Built with open-source models only. No GPT-4, no Cursor, no vendor lock-in.

Created by Vinayak Rastogi.

Architecture

Key Design Decisions

• 3 specialized agents, not 1 mega-prompt. Each phase gets its own agent with a focused prompt and dedicated model. Discovery interviews better when it isn't also thinking about RICE scores.
• Code orchestrator, not LLM orchestrator. Phase transitions are deterministic if/else routing with user-confirmed handoffs. No hallucinated transitions, fully debuggable.
• Mixture of Experts model routing. Three open-source models assigned by task type -- a 20B reasoning model for conversation, a 70B model for spec writing, and an 8B model for extraction/classification. Most turns cost fractions of a cent.
• Open-source models only. Zero vendor lock-in. Swap Groq for Together AI or any LiteLLM-compatible provider by changing one line in config.py.
• 3-layer eval framework. LLM-simulated founders (5 personas), 27 deterministic assertions, and an LLM-as-Judge scoring 5 rubric dimensions. Eval runs end-to-end with one command.
• No frameworks. No LangChain, no LangGraph. Plain Python + Pydantic. The entire orchestrator is ~80 lines of if/else.

How It Works

Phase 1: Discovery

The Discovery Agent conducts a PM-style interview covering 8 aspects: target