Learning by Building: How Launchling Became My AI PM Playground

Home • About • Projects • Blog • Resources • Contact

I didn’t start Launchling because I had a burning startup idea I needed to bring to life. Instead, I started it because I wanted to upskill in AI product management fast, meaningfully, and through hands on learning rather than just theory. I could see that AI was becoming a core part of the modern product toolkit, and that if I was going to lead teams building with LLMs, I needed to walk the walk and not just talk about strategy from the sidelines.

So I set out to build a real product, powered by AI, with real users and real constraints. Not only did I have a lot of fun; but the more I built, the more I realised that Launchling might actually solve a genuine problem for early-stage founders (especially non-technical ones). So I kept going.

Now it’s both:

  • 🧪 A learning lab for deepening my AI PM skills
  • 🚀 A product I hope will help others start small and build big when they’re ready

Here’s what I’ve learned, where I’m still growing, and how Launchling is helping me do both.

✅ What I’ve Learned Through Building Launchling

  1. Prompt Engineering Is UX Design

Every output in Launchling, from startup ideas to tailored plans, is generated by GPT-4. But the success of those outputs depends entirely on how they’re framed.

I’ve learned how prompt design directly intersects with:

  • User onboarding (what do I need to ask to generate something useful?)
  • Tone of voice (how do I sound empowering, not prescriptive?)
  • Ethical defaults (how do I steer away from harmful or exploitative ideas?)
  • Output structure (how do I ensure consistency and readability?)

This isn’t just fiddling with text. It’s product design inside the model; and I think it might be one of the most powerful, under appreciated levers in AI products (at least I’m not sure I had fully understood this before).

  1. Designing Guardrails Is Product Work

When you’re building for beginners, especially non-technical founders, you need to bake in trust, safety, and clarity.

I’ve built Launchling to:

  • Throttle usage
  • Log token costs to Airtable for API control
  • Avoid unethical or risky suggestions through prompt filters
  • Provide clear, inclusive language that reduces overwhelm
  • Handle privacy and consent with GDPR-aligned opt-ins and deletion flows

AI safety isn’t just an enterprise governance problem. It’s a product design challenge — and one I’ve approached head-on.

  1. From No-Code to Real Code: Evolving the Stack as I Grew

Launchling didn’t start with a full codebase and that was by design as I wanted to test some of the no/low-code tools out there, and because I wanted to spend my time on mastering the AI side of things.

I began with a Tally form triggering a Zapier flow, then layered on a Framer site. It let me test the concept quickly, but I soon hit hard limits in terms of logic, responsiveness, and state handling.

So I rebuilt the whole product as a React app, hosted on Vercel, using Airtable for structured user data and GPT-4 for generation.

That rebuild taught me how to build and maintain a scalable and maintainable codebase solo. It’s no longer a prototype. It’s a functioning AI product and I now own the full stack.

4. Evaluating AI Output with More than Gut Instinct

That work has helped me understand:

Rather than relying on manual review or qualitative feedback alone, I’ve already built a script to evaluate the quality and alignment of Launchling’s outputs against structured user input. I wrote more about this here.

  • How embedding similarity and scoring logic can reveal misalignment
  • Where GPT-4’s outputs drift from user intent or framing
  • Why even ‘good-enough’ outputs require a clear quality definition

This seems to be one of the most critical but often overlooked areas in LLM product design - how do you know your model is doing what it should?

🧗 What I’m Still Learning

As much as I’ve learned, there are some big areas I’m deliberately stretching into next.

  1. Systematic Evaluation of AI Output

My current evaluation script is a strong start, but I want to go further by:

  • Adding confidence scores at runtime based on user-plan alignment
  • Exploring token-level coherence metrics and comparative output benchmarking
  • Building systems to analyse not just performance but fairness, consistency, and utility across use cases

This is how I’ll evolve from basic output scoring to robust AI product evaluation frameworks.

  1. Agentic Workflows and Multi-Step Reasoning

Launchling currently uses single-shot generation. But real-world users often need more:

  • Step-by-step refinement or simplification
  • Dynamic follow-up based on feedback
  • Self-critiquing flows (e.g. “Is this plan realistic?” → revise)

These are agentic patterns I want to experiment with inside Launchling but it’s important I don’t do this just because they’re trendy, but because they’re useful. So it might be something I experiment with but don’t ultimately release into production. But this is an opportunity to learn about which agentic patterns genuinely improve utility and clarity for my users.

  1. Retrieval-Augmented Generation (RAG)

All Launchling outputs are currently based on structured user input not external sources. I want to experiment with:

  • A startup case study library for inspiration
  • Contextual retrieval from real tool guides or founder FAQs
  • A vector store that tailors advice based on embedding similarity

This will make outputs more grounded, credible, and helpful; and it’s a great chance to deepen my understanding of RAG pipelines in practice.

4. AI UX Research

So far I’ve focused on building good defaults and clear flows; but I want to explore more structured AI-specific UX research, such as:

  • Expectation mapping for different founder personas
  • Mental model analysis of what users think the AI “understands”
  • Failure case interviews to improve plan resilience and utility

This will sharpen my ability to diagnose AI product gaps and align model behaviour with user expectations.

  1. AI Governance, Risk, and Regulation

I’ve already implemented product-level safeguards, but I want to deepen my fluency in:

  • Upcoming regulation (EU AI Act, UK frameworks)
  • Trade-offs around explainability, safety, and autonomy
  • Responsible AI design patterns beyond just guardrails

This is essential for leadership in this space and something I’m exploring through self-study.

🛠️ Why Launchling Still Serves My Learning Goals

One of the most powerful things about this project is that it’s still useful, even as I grow.

I don’t need a new learning environment, I just need to increase the complexity, sharpen the metrics, and deepen the stack. Some of the next experiments I’m planning:

  • ✅ Confidence scoring using embeddings
  • 🤖 Step-by-step plan refinement agents
  • 🔍 Tool database + retrieval for plan grounding
  • 🧭 User archetype detection to tailor onboarding
  • 🛡️ “Build responsibly” nudges and checklists

Every feature is both a product bet and a learning opportunity.

💡 Final Thought

If you want to learn AI product management, you don’t need a course, a title, or a team. You need a real user problem, a willingness to figure things out as you go, and a system that lets you learn in public.

That’s what Launchling has been for me - a way to test my thinking, sharpen my skills, and grow in a space that’s evolving fast.

It started as a deliberate learning project, it might still become a viable product; either way, it’s been a really effective way to upskill.

🪂 Try it now

Turn your idea into a tiny, tangible prototype — with a little help from Launchling.

👉 Try Launchling – no signup, no jargon, just a gentle push forward.