diegosouzapw AI-Native Product Refounding
In the AI era, product market fit must be constantly "refounded." This framework moves teams away from "blunt instruments" (long roadmaps, rigid PRDs) toward a high-velocity, hands-on approach where the product is shaped by the unique capabilities of evolving models.
Core Principles
- Vibes before Evals: During the divergent "discovery" phase of an AI feature, prioritize "vibe-checking" (open-ended testing) over rigid evaluation benchmarks. Converge on formal evals only once the core "Aha!" moment is found.
- The Hybrid Prototyper: PMs, Engineers, and Designers must collapse silos. A PM must be "technical enough to be dangerous" and a designer must understand LLM tool-calling limits to build realistic UX.
- Greedy Inference: Be "intentionally wasteful" with compute for strategic insights. Spend hundreds of dollars on LLM calls to analyze sales transcripts or user data if it yields one "astute" product insight.
The Refounding Workflow
1. Conduct the "Clean Slate" Audit
Before adding AI to an existing feature, ask: "If I were founding this company/feature from scratch today with current AI capabilities, what would the native experience be?"
- Identify if your current product is a "Lego kit" (useful primitives) or "Legacy weight."
- Determine if the AI should be an assistant (sidebar) or the primary agent (the default interface).
2. Bifurcate into Fast and Slow Thinking
Restructure the team into two distinct modes to prevent infrastructure from slowing down innovation:
- Fast Thinking (The AI Platform Group): Focus on near-weekly shipping. Their goal is "jaw-dropping" value and rapid experimentation.
- Slow Thinking (The Durable Group): Focus on infrastructure, data complexity, and scalability (e.g., high-scale data stores) that cannot be "hacked" together in a week.
3. Implement the "Play" Mandate
To understand what models can actually do, the team must use them "hourly."
- Cancel Meetings: Give the team a full day or week to do nothing but play with new AI products (e.g., Cursor, Runway, NotebookLM).
- Project-Based Learning: Force every PM to build a "weekend project" using AI (e.g., a personalized CRM or an automated researcher) to learn the constraints of code-gen and prompting.
4. Move from PRDs to Interactive Prototypes
AI behavior is non-deterministic; you cannot "word-smith" your way to a great experience in a document.
- Show, Don't Tell: Share Replit links or interactive prototypes instead of slide decks.
- Inspect the "Chain of Thought": When reviewing an AI feature, don't just look at the output. Test "unrealistic" prompts to see where the logic breaks.
Examples
Example 1: Refounding a Search Feature
- Context: An enterprise app has a traditional keyword search.
- Old Approach: Create a roadmap to add semantic search and filters over three months.
- AI-Native Refounding: Create a "Fast Thinking" pod to build a natural language agent that crawls the web and internal data simultaneously. Use "vibe-coding" to test if the agent can answer "Which of my podcast guests have never been asked about their failures?" and iterate daily based on results.
Example 2: Strategic Greedy Inference
- Context: A PM is trying to identify why a certain segment is churning.
- Input: 500 sales call transcripts.
- Application: Use an "LLM Map-Reduce" approach. Break the transcripts into chunks, run LLM calls on each to extract "Product Gaps," then run an aggregation LLM call to synthesize the top 3 strategic shifts.
- Output: A high-fidelity report that would have taken a consultant weeks to produce, delivered in 30 minutes for $150 in API costs.
Common Pitfalls
- The "Check-the-Box" AI Feature: Adding a basic chat sidebar that doesn't utilize the product's unique data. If the AI doesn't manipulate the core primitives of your app, it’s just a wrapper.
- Premature Evals: Setting up complex evaluation pipelines before you've found a "magical" user experience. This constrains creativity and slows down the "Fast Thinking" group.
- Role Silos: Waiting for a designer to finish a Figma file before an engineer tries the prompt. PMs should use tools like v0 or Lovable to build the "vibe" of the UI themselves first.
- Polishing the "Golden Path": Only testing prompts that you know work. You must try to "stump" the AI during development to find the necessary guardrails.