Paweł Huryn: If I had to learn AI product management again, I'd start here

Paweł Huryn runs productcompass.pm and writes on AI product management for an audience of more than 10,000 subscribers. This January 2026 article addresses a specific waste pattern he observed among aspiring AI PMs: months spent on linear algebra, Python courses, and backpropagation when none of that knowledge transfers into shipping better AI products.

The article proposes eight areas that Huryn argues actually matter, organized around what a PM needs to understand at a product level rather than a research level.

The first is basic concepts: how LLMs work, what transformers and embeddings are, what context windows do. Not the mathematics behind them, but the mental model for how the technology behaves under product conditions. The second is prompt engineering, which Huryn calls the most underrated skill in AI product work. He describes prompting as product design for AI — the interface between human intent and model output.

Fine-tuning comes third: when it is useful and when it is not. His position is that fine-tuning is often unnecessary if prompting is done well, and knowing the difference saves significant engineering investment. Fourth is RAG (retrieval-augmented generation) — the mechanism for giving LLMs access to custom data — which he treats as a near-essential capability for any product that needs to operate on proprietary or real-time information.

AI agents and agentic workflows come fifth. Systems that plan, reason, and use tools autonomously are becoming the default architecture for complex AI products, and PMs need a working model of how they function. Sixth is AI prototyping: hands-on building using no-code tools such as Lovable, Supabase, and n8n. The argument here is that this skill compounds quickly once started and reveals gaps in product thinking that reading alone does not expose.

Seventh is knowing the major foundational models — Claude, ChatGPT, Gemini — and understanding their different strengths, pricing structures, and limitations, since model selection is a product decision with significant downstream consequences. Eighth, and the one Huryn identifies as the scarcest skill, is AI evals: building the measurement systems to determine whether an AI product is actually working.

The article argues for action-oriented learning over passive consumption. Huryn suggests two months of hands-on work, with the principle that one hour of building teaches more than a week of reading. The framing is practical: learn the minimum needed to ship, then expand from there.

Useful for any PM moving into AI product work who wants to understand where to focus first, particularly those who have been told they need deep technical knowledge to contribute meaningfully to this space.