How I Built a Gem in 20 Minutes (And What I Use It For Every Week)

Amy Westlake
4 days ago
3 min read

Situation

I wrote the same risk description three times before I sent it.

Not because I didn't know what the risk was. I knew exactly what it was. The feedback kept coming back the same way: too focused on the problem. Not on impact. Not on what this means for the organization.

I knew what they needed. I'd been doing this long enough to know. I just couldn't get my brain out of problem-description mode when I was sitting there at the keyboard, staring at a thing that was on fire.

The framing gap had been following me for months. Not big enough to feel like a crisis. Just persistent enough to be annoying.

The AI Move

I described the problem to Gemini. Not the risk itself - the task. I have a repeated thing I need to do: take a raw description of a risk or issue and turn it into four things. A risk description. The detail. An impact statement. And a mitigation plan. All of it needs to stay outcome-focused. Not problem-focused. Can you help me build a Gem for this?

It asked clarifying questions. What kind of environment? What audience? What does "outcome-focused" look like in practice?

I answered them. It wrote a starting set of instructions.

I ran it on a real risk. Something I'd been circling for twenty minutes. The Gem produced all four sections in under a minute. The language was sharper than what I'd been writing. The impact statement actually said what the impact was.

That was the first twenty minutes.

The Shift

Here's where it got interesting.

I used the Gem at the next steering meeting. The framing landed. Positive feedback. I moved on.

Two weeks later, a different risk came back with a note: the mitigation plan still reads like a list of actions. Not a path to resolution.

I went back to the Gem, not to the output, to the instructions. I told Gemini: the mitigation section needs to feel like a coherent response, not a task list. Stakeholder feedback said it reads like actions without resolution. Here's what they meant. Can you update the instructions to address this?

It revised the instructions. I ran it again. The mitigation section changed.

I've done that four or five times now. A piece of feedback lands. I take it back to the Gem. I ask it to improve its own instructions based on what I learned. The next risk I run through it reflects the improvement.

I'm not editing my prompts. The Gem is editing itself.

The Pattern

A Gem is just a saved set of instructions. You describe how you want the AI to behave - the structure to produce, the tone to use, the distinctions that matter - and every time you open it, it already knows. You don't re-explain yourself. You just start.

What I didn't fully understand when I built this one is that a Gem can also learn.

The loop I was doing manually (gathering feedback, adjusting my approach, trying a different framing) became a loop I could run in minutes. Describe the gap. Give the context. Ask it to update its own instructions.

It takes less time than revising a draft. And it improves the next output, not just the current one.

I'd been thinking of Gems as static tools. Build it once, use it until you don't need it anymore. This one changed that. It's more like something that absorbs what you're learning and keeps getting better the more you use it.

The Implication

Build one.

Start with a repeated task that has a specific output format - a few consistent sections, a particular tone, something you produce more than once. Describe what you want. Answer its clarifying questions. Run it on something real.

Then notice what's missing. Not in that output, in the pattern. After a few uses, you'll have feedback: a stakeholder comment, a piece you keep having to rewrite, a framing that still isn't landing. Take it back. Tell the Gem what you learned. Ask it to revise its own instructions.

That's the whole loop. Twenty minutes to build. Much less than that to train.

The thing that surprised me is that you're not just saving time. You're building something that absorbs what you know and applies it consistently, without you having to remember to apply it every time.

What I'm Testing Next

The risk Gem is still evolving. I have feedback from other stakeholders I haven't worked in yet. This means training it further so "impact" and "outcome" resonate with many stakeholders.

But I've also started looking at my other Gems differently.

The one I use for weekly status updates. The one I use to review the quality of my own AI outputs before I share them. All of them were built once and used since. None of them have been trained the way this one has.

I'm curious what they'd become if they were.