Collaborative Writing with AI: A Live Workflow Breakdown

The 'Open Kitchen' Writing Methodology

Dan Shipper, CEO of Every, employs an 'open kitchen' storytelling approach—a term he coined with ChatGPT—to document the history of OpenAI's Codeex. This method involves writing in public and providing readers with direct access to the source material, interview transcripts, and research notes. By treating the writing process as a collaborative, transparent exercise, he aims to create a more verifiable and immersive reading experience.

Iterative Drafting and Reasoning

Shipper uses ChatGPT’s voice mode to manage the cognitive load of drafting. The process involves:

1. Review and Refinement: He has the model read back existing sections to ensure flow and tone, then issues specific editing instructions (e.g., adding technical nuance or adjusting quote formatting).
2. Strategic Brainstorming: Rather than just drafting, he uses the model to synthesize complex strategic arguments. For instance, they debate why 'coding' was the uniquely correct strategic pivot for OpenAI, landing on the concept of 'closed-loop verification'—where agents can run code, receive machine-checkable feedback, and use test-time compute to iterate toward a solution.
3. Parallel Research: While drafting, Shipper tasks the model with cross-referencing internal interview transcripts to verify claims, such as the origins of the Codeex project and the specific contributions of team members like Tibo Sautio and Alexander.

Managing Complexity and Bottom-Up Innovation

Shipper explores the tension between top-down company goals and bottom-up execution. He argues that OpenAI’s success with Codeex wasn't the result of a single, monolithic plan, but rather a strategic priority that acted as an umbrella for multiple, independent experiments. He works with the model to articulate this 'chaotic' but effective organizational structure, ensuring the narrative reflects that there were 'many beginnings' to the project rather than one clean origin story.

The Role of 'Test-Time Compute'

Central to the narrative is the shift from training-heavy models to reasoning models that utilize 'test-time compute.' Shipper and the model refine the argument that programming is the ideal domain for this paradigm because it provides abundant, verifiable feedback loops. This allows the model to spend more compute on 'trying, checking, and revising' solutions, turning reasoning into measurable improvement rather than just longer output.