Call Me MandyProduct craftsperson. Curious by nature.
Seven panels side by side, fading from photo to sketch. A woman works at a desk with a laptop and two monitors beside a window of autumn trees. A monitor shows a product-surface wireframe. A Trip Wrangler screen titled London Arrivals and Activities. A roadmap timeline. A flowchart. A hand writing in a spiral notebook. A notebook page of handwritten notes and sketched wireframes with a pen.

[3] AI as daily practice

Quicker decisions from earned judgment

AI is the fastest part of my practice now. Research, a working prototype, a first draft of the thing I was going to write anyway. What it doesn’t do is decide what’s worth building.

For years my job was to describe what should be built. I ran discovery sessions, owned roadmaps, led launches and shipped things that mattered to people. The prototyping and building were someone else’s part. Then AI handed me the keys, and I haven’t put them down.

Early on, I was asking an LLM to write product requirements. They were impressively useless. More than a year later, I’m researching, prototyping and building as a daily practice, and the requirements are better for it. Being able to show — not just tell — the nuance of a concept helps me and my teams spot a bad idea before spending time and money on it.

CMM Labs & Consulting

Learning out loud

I build with AI to learn and test ideas. Visit CMM Labs to see what I’m making and learning along the way.

Visit CMM Labs ↗

CMM Labs & Consulting

Learning out loud

I build with AI to learn and test ideas. Visit CMM Labs to see what I’m making and learning along the way.

Visit CMM Labs ↗

AI adds to the craft

My product practice using LLMs

These are habits that daily use and study of LLMs built for me. I continue to explore different tools and methods to evolve with the rapidly moving space.

  1. Research before I decide

    Before I form an opinion about a market, a customer or a competitor, I ask for the landscape first. I do it often enough that it’s a habit rather than a step, and deciding what the evidence means is still mine.

    I wanted to know how in-house talent acquisition screens for AI experience, so I asked 18 of them. All self-selected, all in the first six days.

    Self-selected answers flatter. The useful signal was that recruiters lean on keyword filters they don’t trust, which sent me to 20 live postings to see what competencies actually get named.

  2. Prototype instead of describe

    When a decision won’t settle in conversation, I build the smallest thing someone can click. Daily use is what made that reflexive, and it’s usually faster than arguing about the description.

    A planning model I’d kept in a spreadsheet became a tool for one PM and one planning scenario. The first concept was over-engineered.

    The scope was the problem, not the features. Narrowing to one PM and one scenario is what made it useful, and the judgment stays inspectable.

  3. Evaluate what comes back

    AI writes confidently, including when it’s wrong, so I label a claim before I use it: stated, implied, inferred or unverified, with the source when there is one. Doing it every day is what made it automatic.

    Claude edited an essay for a role I wanted and told me I’d increased revenue on a product by $500,000 in 2025. I hadn’t. That figure came from a role I’d held years earlier.

    The check that caught it was my own memory of my record, which no model has. The labels are standing instructions now, and I still triple check.

Beneath the surface

Behind the hype is the 20%

A brilliant engineer I worked with put it to me plainly. AI is great for 80% of the work. That last 20% is where it falls down.

This engineer spent years inside a platform that had served hundreds of customers, and he knew exactly which 20% he meant.

The platform had thin documentation, few people left who understood the full architecture and hidden dependencies from years of customization without oversight. That’s the 20%. It’s unknowable from the code.

I use these tools daily, and watching an idea become a clickable thing in 10 minutes is a little bit magic.

But the 80% was never what made this work expensive.

An article I read put it this way. Code generation is the easy part, trust generation is the hard part. His answer is evaluation as a real engineering discipline, including whether the model exposes its assumptions or quietly invents them when the legacy system is ambiguous.

You can build a suite that catches an invented assumption. You cannot build one that knows the rule exists because of a 2008 contract, or whether that contract is still in force.

For the systems product teams inherit and steward, AI is a bolt-on. Everyone’s watching the 80%. The 20% is the part nobody can see.