Karina Nguyen on Writing as AI Development Practice

Table of content

Karina Nguyen builds AI systems at the frontier. She contributed to Claude 3 at Anthropic and now works on ChatGPT Tasks, Canvas, and the o-series models at OpenAI. But her most interesting insight isn’t about model architecture—it’s about how writing shapes research itself.

Writing as Research Infrastructure

At Anthropic, Nguyen developed a practice that many AI researchers overlook: extensive daily documentation.

“I personally started writing an extensive doc on my daily research progress and thoughts we have during meetings to capture the evolution of our approach to hard problems. I also log all the bugs I encounter and how they were resolved as a way to get better at programming.”

This isn’t just note-taking. It’s active thinking.

“This way I learn where analogies break down, and where new metaphors are needed. I discover gaps in my own knowledge and often this regular practice of refining fuzzy thinking into precise expression sharpens my mental model on how RL works.”

The Slack Channel Method

One cultural pattern Nguyen observed at Anthropic: personal Slack channels for work-in-progress.

“Another thing I love at Anthropic is that we each have our own Slack channels to share works in progress and thoughts. People can subscribe to these channels. I feel very grounded to have a place where I can share not fully formed thoughts or questions.”

This creates space for half-baked ideas without the pressure of formal documentation. Colleagues opt in, criticism flows naturally, and thinking improves through exposure.

Prompting as Writing Practice

For Nguyen, even the daily work of prompting models becomes a form of structured writing:

“Prompting models becomes a form of writing that every research engineer engages in daily. This kind of writing requires forming hypotheses, testing assumptions, and revising based on new insights.”

The feedback loop is immediate. You write, the model responds, you refine. Each iteration tightens your thinking.

Core Lessons from Hypergrowth

Nguyen joined Anthropic at 50 people and left at 700+. Her takeaways from that scale:

“The pace of a team’s progress is largely a function of its decisiveness and open-mindedness to take risky bets”
“The simplest and dumbest approach will often just work”
“You have to go through the entire journey of full understanding to arrive to the simplest answer”
“Every time you train a new model there will be an inevitable brain damage that needs to be solved”

That last one is instructive—model issues can often be reverse-engineered by carefully examining training data. Pattern recognition comes from hands-on documentation.

Beyond AI: Investigative Roots

Before AI research, Nguyen worked on visual forensics and human rights investigations at the New York Times, contributing to Pulitzer Prize-winning reporting. Her work included war crimes documentation, satellite analysis, and 3D reconstructions.

This background shows up in her approach to AI work: meticulous evidence gathering, verification processes, and systematic documentation. Research is investigation.

The Writing Culture Thesis

Nguyen’s central argument: organizations that treat writing as creative practice—not bureaucratic necessity—produce better work.

“If innovation thrives on imagination then organizations must cultivate an environment where writing is treated as an act to give ideas the space to grow and resonate.”

Strong writing cultures share common traits:

Tools that reduce friction for expression
Permission to write rough drafts publicly
Iterative processes that refine thinking
Recognition that early efforts will be ragged