Designing AI Products: The Learning Path NN/g Actually Recommends

Most product designers get handed an AI feature the same way: a roadmap slide, a deadline, and no method. NN/g’s Designing AI Products and Features study guide, published September 18, 2026, is the closest thing to a curriculum for that problem. It’s a link collection, not an article, which is exactly why it’s hard to use — nobody tells you what order to read it in.

Here’s the order that makes sense, and the parts you can skip.

Start with the question that kills half of designing AI products

NN/g’s framing is blunt: adding AI doesn’t create value by itself, and in some cases it’s a large investment for a small return. The guide points back to Jakob Nielsen’s 2021 warning about overinvesting in new tech before understanding the user need — written before ChatGPT existed, which is the part worth noticing.

Read the value questions before you read anything about interface patterns. If your team can’t answer why the AI version beats the non-AI version, no amount of prompt-control design will save the feature.

The guide names four things AI is actually good at: content creation, summarization, basic data analysis, and perspective taking. That’s a short list on purpose. If your feature isn’t doing one of those four, that’s the moment to ask why you’re building it.

The value gate: three questions the guide points to

NN/g links a piece built on advice from Condens‘ cofounder with three questions for judging whether an AI integration is valuable and appropriate. The guide doesn’t spell them out in the index, so treat this as the section to read in full rather than skim — it’s the decision gate for the rest of the project.

Two more findings from the same collection are worth carrying into that conversation:

• Narrow beats broad. Tightly scoped AI features are easier to understand and get better adoption than sprawling ones.
• Novelty doesn’t pay. AI built for novelty rarely produces real value; the guide’s advice is to solve a real pain point instead.

There’s also a warning about chat. The guide says teams should avoid rushing to chat interfaces unless they meet real user needs — and separately notes that users see little reason to use a site-specific chatbot unless it does something the existing site can’t.

The prompting problem is a design problem

This is the most useful section for anyone shipping this quarter. NN/g’s position is that the power of LLM features is largely mediated by whether users can write a good prompt — and users often don’t know what they want or what the model can do.

Prompt support is interface work, not documentation work. The guide’s recommendations:

• Hybrid interfaces that pair prompt input with a graphical UI improve usability for image generation.
• Well-designed use-case prompt suggestions help learnability and set realistic expectations.
• Suggestions need to be contextually relevant, personalized, and matched to the user’s experience level.
• Prompt controls have four jobs: discoverability, education and inspiration, setting constraints, and enabling follow-ups.

If you’re designing an AI feature right now, that list is your spec for the empty state. Most teams ship a bare text box and call it done.

What the guide says about writing and visibility

Two smaller findings that get ignored in practice. First, generative output still has to follow web-writing rules — concise, scannable, inverted pyramid, plain language. Second, even valuable AI features have low impact if nobody notices them; the guide uses Amazon’s Rufus as the example.

There’s also a note on the sparkles icon being inherently ambiguous, which is the kind of detail that tells you the research came from watching real sessions rather than theorizing.

Where this guide stops

It’s a link collection. There’s no single worked example, no template, and no pricing or tooling advice — you won’t find a recommendation for which prototyping tool to use. If you want a checklist you can hand to a PM, you’ll have to build it from the pieces above.

The other gap is measurement. The guide is strong on deciding whether to build and on interface patterns, and thinner on what to instrument after launch.

How designing AI products fits the tools you already use

NN/g stays tool-agnostic, so the practical question is which of your existing tools absorbs the work. The three below are the ones most product teams already pay for, and each one carries a different part of the AI design job.

Read the table as a division of labour, not a ranking. None of these tools answers the value question for you — that’s the point of the guide’s first section.

If you’re a freelancer scoping AI work for a client, the value questions are also your scope document. A client who can’t answer why the AI version is better is a client who’ll ask for revisions on a feature that shouldn’t exist.

What to do this week

Read the value section first, then the prompting section. Skip the agent-context material unless you’re building agentic features — it’s a different problem with a different audience.

The one thing to take away: the guide’s own framing is that AI is a hammer, and the failure mode is hunting for a nail to hit with it. Lead with the value the feature delivers, not with the technology that delivers it.

Sources

• Designing AI Products and Features: Study Guide — NN/g, September 18, 2026