This post originally appeared on Tommy Keely's personal blog.

At Amplitude, everyone is considered a builder: not just engineers, but PMs, data analysts, designers, and marketers. It’s exciting and energizing, but can also create a real tension. For example, how do you move fast while maintaining quality? How do you know if the things you’re working on are the most important? How do you make sure one person’s breakthrough doesn’t stay siloed, but rather immediately uplevels the entire team?

These are questions we constantly wrestle with at Amplitude. Over the past few months, our teams have been building and refining a set of AI skills that we could use ourselves. They include things like writing PRDs, designing experiments, decomposing metrics, synthesizing customer research, planning launches, and more. We built them because we needed them, and they’ve helped us weather this transition to everyone becoming a builder.

Two months ago, we open-sourced them in a repo called Builder Skills because we think everyone building with AI deserves a shared, strong foundation. It’s free to use, fork, and contribute to.

Since then, our repo’s gotten over 100 stars and 15 forks just through organic interest. We wanted to post about it to broaden the message.

→ github.com/amplitude/builder-skills

What’s in the library

The library covers the full builder stack, organized into five areas:

• product-skills for when you’re staring at a blank doc, trying to transform a half-baked idea into something you or your team can actually build. These skills are for structuring PRDs, getting the most out of customer discovery sessions, designing strong experiments, and reading out results to actually drive a decision. The skills are built around proven frameworks so the AI isn’t just generating a solution or output, it’s helping you work through the why behind the problem.
• analytics-skills are great for when you have data but aren’t sure what it’s telling you. Sure, these skills will help with building charts and dashboards from plain language. But more importantly, these will drive a diagnosis of things like why retention is flat, whether or not a finished experiment should GA, and synthesize a mountain of customer feedback into something prioritizable.
• growth-skills can help you figure out where to focus. The metric tree skill alone is worth the download: it forces you to decompose a top-line number into its actual components, size every node with real math, and just as importantly call out what areas might be tempting but the leverage isn’t actually there. I’ve found these skills to be particularly helpful for grounding your team in your data & avoiding emotional prioritization.
• execution-skills for the operational work that quietly eats your week. You know what you should be doing, but somehow half your time disappears into summarizing meetings, writing updates to get alignment, and documenting things that already happened. These skills handle that layer so your energy goes toward higher leverage work.
• launch-skills because shipping something is only half the battle. The work doesn’t count if no one sees it. These skills cover the full motion from strategy and messaging to blog posts, social copy, landing pages, and distribution — so the thing you spent weeks building actually reaches the people it was built for.

I don’t know what a ‘skill’ is and at this point I am too scared to ask ...

Maybe you’re reading the above and thinking, this all sounds great, but what does it all actually mean ...

You’ve already been prompting AI for years. A skill is just a prompt that’s been done right: it’s a structured, repeatable template that tells your AI not just what to do, but how to do it. A skill is the right framework, the right sequence of questions, the right output format for a specific task. You bring your context (your notes, your data, your half-formed idea), and the skill handles the rest.

The difference between a prompt and a skill matters. Asking an AI to “help me write a PRD” will get you somewhere. A skill gets you somewhere good, consistently. It’s the difference between telling a new hire “write a spec” vs. walking them through how your best PMs actually do it. Detailing what goes in, what gets cut, what questions need answering before a single word gets written, etc.

Skills work across any LLM: Claude, ChatGPT, Cursor, whatever you’re already using!

Why “battle-tested” is the thing that matters

There are a lot of AI skills repos out there right now. Most of them were written by someone who has never actually used the skills collaboratively inside a real company, on real work, under real constraints.

These skills were built and refined by the builders at Amplitude actually doing product and growth work. The build-metric-tree skill, for instance, came directly from the kind of metric decomposition work our growth PMs do to identify leverage points and avoid distractions. It’s not theory, it’s something we’ve wrestled with and collectively encoded so we reduce repeated mistakes and amplify wins.

That’s the standard we’re holding the library to: skills that have earned their place through actual use, not skills that look good in a demo.

How to get started in 3 steps

1. Install as a plugin If you’re using Claude Code or Claude Cowork, you can install the entire repo or individual discipline folders as a plugin. Once installed, skills and commands appear in your skill selector and are available to trigger by name.
2. Use as prompt templates No special setup required. Browse any plugin’s skills/ folder, open the SKILL.md, copy the prompt template, paste it into your LLM of choice, fill in the {{PLACEHOLDERS}} with your actual context, and run it. That’s it.
3. Pick your first skill If you’re not sure where to start:

• Overwhelmed by a messy product problem → craft-spec
• Not sure where to focus for growth → build-metric-tree
• About to do customer discovery → mom-test
• Experiment just finished → craft-experiment-readout
• Planning a launch → launch-strategy

We’re publishing this because we think the future of AI-native work is built on shared, composable primitives, not proprietary prompts locked in someone’s private folder. The more people use these skills, improve them, and add their own, the better the whole library gets for everyone.

If a skill doesn’t work for your context, open a PR. If there’s a framework you use that isn’t in the library, add it. If you build something interesting on top of this, we want to know about it!