Great documentation with Han Wang from Mintlify
Mintlify co-founder Han Wang explains why eight pivots made product-market fit unmistakable, why documentation quality is mostly an operating habit, and how AI turns readers’ questions into a new feedback loop for the product team.
Han Wang and his co-founder Hanbi built products together at Cornell before leaving big-tech jobs to start a developer-tools company. Their first viral AI coding product faded quickly, and seven more ideas failed before Mintlify found demand.
Han tells the story of turning one friend’s documentation into the first customer, then explains how teams can treat docs as a product: complete, current, easy to change and increasingly responsive to the exact questions users ask.
The Q&A below is experimental — extracted from the episode transcript with AI. If you spot a mistake, let us know at jack@scalingdevtools.com.
How did Mintlify emerge from the original startup?
We began with an early Codex-powered tool that translated and explained code. It attracted a large launch spike but steadily decayed because it did not solve a durable problem. YC still backed us, and we kept experimenting within developer communication. It took roughly eight pivots and repeated customer interviews before complaints about bad documentation finally became the problem we chose to tackle.
What did the first recognizable version of Mintlify look like?
Over a weekend we rebuilt a roommate’s weak docs with polished design, search, code snippets and Markdown in the repository, then helped switch the DNS. His team loved it. A public post brought around ten customers in the first week and roughly twice that by the next, with people paying immediately. After so many failed ideas, overflowing demo demand felt unmistakably different.
Why did Paul Graham tell you to change the name?
Months into the business, Paul invited us to his home. We introduced ourselves and he immediately called Mintlify a terrible name, then spent the opening minutes trying to persuade us to replace it. We kept it. The story is funny in retrospect because the name he disliked became increasingly familiar across developer tools.
What is the minimum standard for good documentation?
A reader should find the correct, current information and complete the job quickly. Beautiful presentation helps, but the first obligation is coverage: common questions, product facets and operating steps must actually be written down without contradictory versions. Documentation becomes harmful when it subtracts from an otherwise good product by blocking the person trying to use it.
How do teams maintain that standard?
The answer is less a page template than a product process and culture. Define the audience, put the necessary information on paper, collect feedback and iterate. If the docs vanished, many developer products would effectively disappear too, so teams should manage them like a core product surface rather than collateral completed after the feature ships.
How can developers be encouraged to contribute?
Do not make excellent prose a prerequisite for recording useful knowledge. Reduce the friction for people who may dislike writing by keeping docs in Markdown alongside code. They can update a page in the feature pull request, preview it through the familiar review flow and publish by merging to main. The workflow turns documentation into part of engineering rather than a separate ceremony.
Where should tutorials, guides and references live?
Teams often worry about perfect information architecture before they have written the information. Missing content is usually the larger failure. Put the answer somewhere, make it discoverable, then improve how guides, quick starts and references connect as the corpus grows. Search already helps readers enter at the relevant page instead of following the designed path from the homepage.
How is AI changing the way people read docs?
Traditional docs resemble a tree: the reader navigates from a homepage through branches toward one answer or a sequence of concepts. A model can jump directly to the answer grounded in that tree. The source documentation still has to exist and remain accurate, while curated pages continue helping people discover what they do not yet know to ask.
How can AI improve the documentation itself?
Chat reveals full questions instead of only page views, search keywords or a vague thumbs-down. Teams can compare those questions with the generated answers, see where the corpus is missing or misleading and repair it each week. As model errors decline, a wrong answer increasingly exposes a content gap — and sometimes even a product capability users expected but cannot find.