How I AI: Marily Nika’s AI Native PM Workflow (Perplexity, Veo, v0, Notebook LM)

Marily Nika, an AI Product Lead at Google and founder of the AI Product Academy, demonstrated a rapid product-concept workflow that moved from desk research to a PRD, prototype, and concept video in about 20 minutes.

Her example began with a smart fridge notification saying a Coca-Cola had been stored for 80 days. She used that odd moment as the starting point for a broader smart-fridge concept.

In this episode of How I AI, Marily uses Perplexity to summarize public Reddit discussions, a custom GPT to draft a PRD, v0 to create a clickable prototype, and Flow and Sora to explore a concept video. She also demonstrates NotebookLM as an additional voice during her bootcamp demo day.

Marily calls the approach tool hopping: use a specialized tool for each stage instead of forcing one model to handle the entire workflow.

Move from an idea to a concept package in 20 minutes

How to Go from Product Idea to Prototype in 20 Minutes with an AI-Enhanced PM Workflow: Move a product idea through evidence gathering, a structured opposing-view debate, a decision-ready PRD, a clickable prototype, and a short vision asset without confusing generated output for validated demand.

The demonstration connects rapid desk research, a draft requirements document, an interactive prototype, and a vision video. It accelerates exploration but does not replace customer interviews, validation, or production design.

Scan public discussions with Perplexity

Marily begins with a fast scan of public discussions to identify objections and language worth investigating. In the demo, this took about three minutes.

Tool: Perplexity.

She selects the Discussions focus to emphasize forum sources such as Reddit.

Initial query:

would families be interested in a smart fridge?

She then asks two simulated agents to argue for and against the idea. This is a way to surface objections from the collected material, not evidence that the model represents actual users.

Create two agents, one that is pro smart fridge and one that is against smart fridge. Use everything you read and have these two agents debate at least, I dunno, like 20 times about it, and give me the minimum set of features I would need in order to convince the against agent.

Perplexity returns a debate and a proposed minimum feature set. Marily carries ideas such as local processing for privacy into the next step, where they still require validation.

Draft a PRD with a custom GPT

Marily pastes the proposed feature set into a custom GPT configured with her preferred PRD structure and writing style.

Tool: a custom GPT with her PRD template and style.

Process: move the feature set from Perplexity into the custom GPT.

Instruction:

generate a PRD about a smart fridge that has these features

The custom GPT produced a draft PRD in about 90 seconds, including a problem statement, target users, feature analysis, and prioritization. Marily treats it as a starting point for product judgment, not a finished specification.

Turn the PRD into a clickable prototype with v0

Marily uses the draft PRD as input for an interface that stakeholders can inspect and discuss.

Tool: v0.

Process: copy the generated PRD into v0.

Instruction:

Create the UI of a smart fridge, given this PRD as an input

v0 generated a smart-fridge dashboard with temperature, power, door status, local-processing language, and diagnostics. The prototype made the concept concrete enough for review while remaining a generated exploration.

Explore the product story with generated video

Marily then created short concept videos to show the proposed product in use.

Tools: Google Flow and Sora.

Flow instruction:

create a promotional clip for two, for a couple, using a smart fridge that has these features

The first Flow generations had continuity errors, including an implausible screen placement and a person changing identity. She revised the approach rather than treating the first output as usable.

In Sora, Marily used Cameo to create another concept clip featuring herself and Mark Cuban. The result included dialogue and a personalized reference to Greek salad. Teams using recognizable likenesses need permission for that use.

The video joined the PRD and prototype as another artifact for discussing the concept. It illustrated the intended experience but did not validate customer demand or technical feasibility.

Add NotebookLM to a demo-day judging panel

How to Use NotebookLM as an AI Judge for Product Pitches and Competitions: Use NotebookLM as a consistent additional judge by loading every pitch separately, applying an anchored rubric, and generating an evidence-linked ranking that complements human judges.

Marily runs an AI Product Management Bootcamp with a demo day. She uses NotebookLM as an additional perspective alongside the human judges, not as an impartial authority.

Tool: NotebookLM.

Marily uploads each pitch recording to NotebookLM as a separate source. Anyone adopting this pattern should establish participant notice, permission, retention, and access rules before uploading recordings.

Marily asks NotebookLM to identify three projects using stated criteria:

• Innovation.
• Impact.
• Storytelling.

NotebookLM can generate an Audio Overview of the pitches. Its interactive mode lets Marily ask the generated hosts to announce a selection during the event.

The AI result becomes part of the demo-day experience and can be compared with the human judges. It does not remove model bias or determine the official outcome on its own.

What tool hopping changes

The demonstration shows how quickly a PM can create artifacts for discussion. Research quality, customer evidence, product judgment, and feasibility work still determine whether the concept should proceed.

Tool hopping lets Marily use different systems for search, writing, interface generation, video, and synthesis. Human review connects the outputs and catches the errors between stages.

Marily expects practical AI fluency to become part of the product-management role, especially the ability to choose tools, connect outputs, and review their limitations.

A practical way to test the workflow is to choose one product idea, time-box the exploration, and label every generated output as a hypothesis until it is checked with customers and the delivery team.

Conduct Instant User Research with Perplexity Research whether [target user] experiences [problem] and would value [idea]. Use public discussion sources from [date range]. Return recurring needs, objections, alternatives, counterexamples, and representative links. Then have a pro and skeptical analyst debate the evidence. End with the smallest feature hypothesis, riskiest assumptions, and direct research questions. Do not present the debate as customer validation.

Generate a Product Requirements Document (PRD) with a Custom GPT Draft a PRD from this evidence packet. Preserve source links. Separate observed evidence from hypotheses. Include target user, problem, user narrative, goals and measurable success, critical path, requirements, non-goals, privacy and technical constraints, decisions, open questions, and a validation plan. Do not invent market size or customer claims.

Create a Clickable UI Prototype with v0.dev Create a clickable prototype for the critical path in this PRD. Use realistic synthetic data, make assumptions visible, preserve the stated privacy constraints, and do not add unrelated features. Include the states needed to test [risky interaction] in a product review.

Move a product idea through evidence gathering, a structured opposing-view debate, a decision-ready PRD, a clickable prototype, and a short vision asset without confusing generated output for validated demand.

Use NotebookLM as a consistent additional judge by loading every pitch separately, applying an anchored rubric, and generating an evidence-linked ranking that complements human judges.