As a product designer, I've integrated LLMs into my workflow to amplify productivity

They're like having a team of virtual assistants for tasks like ideating features, writing technical specifications, & refining go-to-market strategies

Here’s how I make it work

One of my key tools is OpenRouter, which lets me leverage multiple LLMs, each with their own strengths (& weaknesses):

- ChatGPT: Ideal for long-form, detailed brainstorming
- Grok: Great for short, focused, & to-the-point responses
- Claude: Perfect for technical or complex questions

Each model serves a unique role in my workflow

The real power of this setup lies in tailored system prompts

I use OpenRouter to create dedicated chat environments for specific tasks, like:

- Brainstorming product features & user flows
- Writing technical or user-facing documentation
- Ideating go-to-market strategies

Segmentation is key to focused outputs

One trick: I maintain a "system prompt chat" specifically for crafting & testing system prompts

By iteratively refining these prompts, I ensure that my task-specific chats deliver accurate & high-quality results

I’ve also learned that structured prompts dramatically improve the quality of responses

Two frameworks I often use (among others):

- TAG: Task, Action, Goal
- RTF: Role, Task, Format

These frameworks force clarity & ensure precise, actionable outputs

Here’s how I might use the TAG framework for improving user onboarding:

- T (Task): Identify pain points in the current onboarding process for new users
- A (Action): Act as a first-time user unfamiliar with the product
- G (Goal): Highlight areas of confusion & suggest ways to make the process faster & more intuitive

By structuring prompts this way, I can pinpoint friction points & get actionable recommendations for enhancing the user experience

For planning a usability test, I might use the RTF framework like this:

- R (Role): You’re a UX researcher preparing a usability test for a new feature
- T (Task): Create usability test scenarios that reveal potential friction points for first-time users
- F (Format): Organize responses by scenario objective, user actions to test, & key metrics to evaluate success

This approach provides a structured & actionable plan for uncovering UX improvements

When using LLMs, I think of them as thought partners, not decision-makers

For example, when ideating design solutions, I might ask:

- What are three alternative user flows for this feature?
- What edge cases might cause this process to break?
- How can we make this experience more intuitive?

These back-&-forths help me refine ideas faster

Ultimately, I’m the curator of what the LLM produces

I own the responsibility for the final output -- whether it’s a user flow, a design specification, or a set of edge cases

No matter how sophisticated the LLM is, the burden of judgment is still mine

This process didn’t come to me immediately

I learned through trial & error -- overcrafting prompts, underexplaining context, & producing mediocre results before figuring it out

My advice?

Experiment boldly & refine continuously

Lessons learned from using LLMs effectively as a product designer:

- Informed prompts produce quality results
- Bad outputs are usually caused by insufficient context
- Refining through iteration & patience is part of the process
- You’re still the source of truth -- don’t outsource judgment

For product designers, LLMs can be transformative

They allow you to brainstorm, validate, & execute faster than ever

By using structured frameworks & refining workflows, you can tap into their full potential

Using LLMs effectively is an iterative process

I rarely accept the first response at face value

Instead, here’s how I refine:

- Provide sufficient context upfront
- Ask follow-up questions to clarify or expand
- Experiment through trial & error, tweaking the prompt until the response aligns with my goals

Context is everything

Most poor outputs stem from insufficient starting information

For example, if I’m writing a feature specification, I'll include:

- The problem we’re solving
- Goals of the feature
- User personas
- Any constraints or edge cases

This improves the output dramatically

That said, overloading LLMs with too much info can lead to hallucinations

The sweet spot is to balance detail with focus: provide essential context while keeping the task clear & concise