Lessons from a Coinbase PM: AI Docs & Mintlify

At ETH Denver this year, I asked a few developers to try creating a wallet programmatically starting from our CDP docs homepage — every single one of them struggled immensely.

Developer documentation is the lifeline for both developers and AI. If people can’t follow your docs, they can’t use your product (and neither can AI applications). As investor Gokul says, “truly great software companies are self-serve first” . This is even more true in the developer space where developers just want to get their hands dirty and go from signup to first API call without having to talk to sales or support.

To address these findings, we partnered with Mintlify to completely uplevel our docs experience from the inside out. It was one of the coolest launches I’ve gotten to lead as a PM at Coinbase.

Here’s what we learned.

Photo from developer experience interviews we did at ETHDenver (that’s me in the brown hat!)

1. Information architecture matters

One problem I noticed immediately when watching people try to navigate our docs, was how quickly & often they’d get lost. Every doc site structures its content differently (tabs, sidebars, dropdowns, switchers etc.). No solution is perfect, but we had to find a structure that better suited our products. 

A few examples:

1. Problem: API reference pages were hard to find. API references are crucial when diving in & testing endpoints, but our API reference pages were split up & nested within each different areas making them hard to find.  

   1. Solution → We consolidated all API references & endpoints that made sense to be used together under a single tab.

2. Problem: Products were difficult to navigate between. We had vague categories and no clear starting point for people to anchor on so we noticed people would click around aimlessly a lot unsure where to find things.

   1. Solution → We renamed our categories and persisted a single left nav on the home page so that users know exactly where to get started and can easily follow from top to bottom.

3. Problem: Search results were disorienting. Our search often provided inaccurate results and sometimes didn’t have breadcrumbs so users wouldn't know where they’d be dropped into.

   1. Solution → We worked with Mintlify to improve our search accuracy and ensure all results had clear breadcrumb navigation, making it easy to understand context and backtrack.

2. AI matters 

One developer at the hackathon asked Cursor to help them create a CDP wallet by copying over a link to our docs. Pretty quickly, though, Cursor got lost trying to parse through our docs. It became clear that our docs were not set up at all for AI tools & vibe-coders.

AI vibe coding isn’t a fad, developers are increasingly pairing with AI tools to code faster. Docs will soon need to be able to power support agents, coding assistants, and autonomous workflows. Our docs need to be knowledge systems that can serve humans and machines as first-class citizens. That’s why we leaned hard into launching AI first features.

A few examples:

1. Easy export to AI tools → We have capabilities now to instantly copy the markdown of any page into ChatGPT, Claude, or Cursor to ask questions, generate code, and explore usage—all without jumping tabs. Perfect for fast iteration and AI-paired development.

2. Quick MCP server setup →  We launched an MCP server you can easily connect to AI tools to allow them to directly query all CDP APIs, docs, and references in real time.

3. Built-in AI Assistant → We launched an AI assistant natively that can provide contextual answers and guidance directly in the docs—triggered right from code blocks.

4. AI-native formatting → We expose structured files like /llms.txt and /llms-full.txt so large language models can read and index the docs accurately for better AI SEO.

Copy markdown feature

Copy markdown feature in action

3. Content & internal efficiency matters

At the hackathon, a couple users got stuck on a guide step. When we checked, we discovered the content was outdated. Fixing it took nearly an hour.

Great information architecture & AI features are important, but neither matters if the content is low quality or just wrong. As a developer, it’s the worst feeling being stuck on something for hours, only to realize the content of docs you were looking at were just simply outdated or misleading.

Looking internally, we realized part of the issue was that our internal workflows made content iteration and updating painfully slow.

Here’s what we changed:

• 50x faster preview link build times  

  • Problem: Previously, building a preview link for any docs change took ~30 mins, making it extremely tedious for people to collaborate together on docs changes

  • Solution: Working with Mintlify we were able to bring building a preview link down to <1 min, making docs iteration 50x faster

• Autogenerated preview links per PR  

  • Problem: Previously also, to create a preview link, people had to manually deploy to one of 5 shared preview links meaning they also might deploy over each other

  • Solution: Working with Mintlify we were able to enable a preview link to be automatically generated per PR 

4. Working with Mintlify

When we first considered revamping our docs, we knew the lift would be heavy. During a brainstorming session, though, a couple people on the team came across Mintlify, an AI-powered documentation platform. We did some research and quickly found out that we could achieve a lot of what we wanted a lot faster and with higher quality by working with them as our docs platform provider.

We reached out to their team, and shortly after signed a deal. This bottoms up approach helped us move like a startup, we launched our new docs site with Mintlify in just 6 weeks.

An added bonus: by investing in Mintlify’s platform, we now inherit the benefits of any new features they release—like better MCP server support or automatic language translations—at no extra cost.

Mintlify & Coinbase partnership announcement

Closing thoughts

Launching an entirely new docs platform has taught me a lot about how important being customer-focused is when shipping delightful developer experiences, and reminded me that you can always get creative to execute efficiently & sustainably. 

Specifically:

• Docs are a product, not an afterthought

• AI is now part of the dev loop. Docs must be machine-readable and AI-friendly

• Speed matters. Fast internal workflows are essential for keeping content fresh and accurate

After we shipped our new CDP Docs site, I took another stab at testing out my prompt from ETH Denver, only this time with Cursor. I used the copy markdown feature and asked it to create a wallet for me – this time the result was much better :) 

Special shoutout to Andrew Kowalczyk, Tiffany Wong, Kevin Leffew, Rose Chang, Anna Engle, Jacky Huang, Josh Nickerson, Han Wang, Ruhan Ponnada