AI Product Coach, Do Product Work With a CPO Coach

I designed a web-based AI product coach, modeled on how a chief product officer could help a non-product manager. It uses Gemini and Anthropic frontier models, using a declarative prompt that forms its coaching identity. It is an interactive app, so it is optimized to keep it fast and affordable.

The problem

Within small organizations, people often get handed product management work, or have to do product management work without even realizing. The complication is that there is no one to teach them or guide them.

So the consequence is that they have to learn it on their own, do research on the web, or try a general chatbot with less than ideal results. They still have to design products, or create strategy, or generate a document artifact that conveys this to others.

What I built

I built a coach that uses Gemini and Anthropic frontier models, modeled on how a chief product officer would coach a new product manager. It asks questions to understand what your goals are, about the ideas or market, and coaches you and teases out the key information you need to build a usable artifact, whether that is a strategy, feature cards, user journeys, etc.

Four intents cover most of what a product manager does while building a product or strategy.

• Build a strategy or roadmap.
• Define a product or feature.
• Learn how to vet ideas with customers.
• Generate a prototyping prompt you can use with AI tools.

It is encoded with several product management frameworks that guide the user on how to build a strategy or a product. For example, some frameworks encoded are SPADE to force a decision, the Rumelt kernel to validate strategy, and North Star to find metrics that matter.

These are just a few, but this gives the model grounding it can use based on real techniques and approaches used in the industry and provides structure. This helps prevent hallucinations and makes it more effective.

How it works

I wrote the coaching as declarative prompting, so it runs with one identity, what I call the Soul Prompt, rather than putting a whole bunch of instructions like traditional prompting.

In the Soul Prompt I defined who it is, what it believes, how it works, what it knows, how it sounds, the tools it uses, and its hard limits. I did this to take advantage of frontier model reasoning capabilities and how they interact differently from older models, which required a lot of instruction and rules.

Frontier models no longer need detailed instructions and rules, as that confuses them, so this approach takes advantage of the core improvements of models over the years. The real benefit is that it can do more and reason better using smaller prompts.

I also created a visual language for the model to use. Rather than giving it every possible artifact it can create, I gave it a visual language it can use to build custom artifacts for product management on its own, still within the product management language, so it can address any type of situation the user needs. You get an artifact whether that is a document, a graph, or a table, with the key attributes included to help you convey your ideas.

Given it is a chatbot, cost matters as each interaction can increase token usage. The use of declarative prompting keeps the prompt small yet very functional. I used caching techniques and optimized every token used, so it can maintain long conversations, keep context, and reduce the overall cost per interaction. A typical complete interaction of multiple turns costs under fifty cents.

Unlike using a generic chatbot and getting information it thinks you need, and hallucinating, this one guides you, walks you through it, and teaches you.

At the end of the day, your information gets translated into something meaningful you can use.

See it

The home screen. Pick one of the four intents and walk through it with the coach.

In practice

“I want a 3 year vision for my company but do not know where to start.”

“My boss asked me to build a dashboard for the sales team.”

“I have an idea for a tool that helps freelancers chase unpaid invoices, but I do not know if anyone would pay for it.”

“I want to prototype a booking app for dog groomers, but I do not know where to start.”

Back to Portfolio