OpenAI Product Lead: Long Documents Are Losing Value in the Age of AI

Whether you like it or not, the 50-page document you stayed up late to write is rapidly losing value.

It’s not your fault—blame today’s AI for being too capable; it can spit out a replacement in minutes that’s just as lengthy, equally well-structured, and nearly flawless.

All workers face this harsh reality:

Today, being able to write a long document doesn't prove you've thought deeply.

These words were spoken by Tara Seshan, Head of Product for OpenAI Codex and ChatGPT Work.

Tara Seshan

In an 81-minute interview on Lenny's Podcast, she also made an even bolder claim:

Building a product for the current model will fail; building a product for the model you imagine one year from now will also fail.

When everyone holds top-tier tools and AI is evolving into a "permanently online agent colleague," where is the competitive advantage for workers?

The logic shared by Tara, which has been successfully implemented within OpenAI, is worth serious reading by every worker who doesn't want to be left behind.

Tara divided the evolution of AI products into three stages:

• Phase One: Chat — you ask, it answers.
• Phase 2: Co-pilot, assisting you in specific scenarios (such as writing code).
• Phase Three: Persistent AI Colleague (Agent), always online, retains context, and independently handles complex tasks.

The boundary between human and AI responsibilities is rising at a visibly rapid pace.

From previously completing the next line of code, to now directly submitting objectives, with automated workflows handling the backend processes.

Whether it’s the “Agent Command Center” on the Codex desktop or ChatGPT Work, which comprehensively covers spreadsheets, reports, and business analytics, both point to the same trend: execution-level tasks are being fully taken over by AI.

In Tara’s words: The agent rows the boat, and humans steer it.

Tara said in the interview that the agent is responsible for rowing, while the human’s role is to steer the boat in the right direction.

When the work of rowing is taken over by AI, the core value left for humans is only five things:

• Set direction (where to go)
• Take responsibility for the results
• Evaluate quality (maintain high standards)
• Inject position and expression (to the soul)
• Collaborate, care for, and motivate the team (win hearts)

Why is it a dead end to say “focus on the present” or “bet on what happens in a year”?

Stay current: Models are evolving too quickly.

You went to great lengths to design a complex product mechanism to work around today’s model limitations—only for all those mechanisms to become worthless “appendices” the moment the model is upgraded tomorrow.

Betting for a year: equivalent to a blind test.

Basing the entire product’s survival on a set of assumptions that are highly likely to fail is destined to lead to total failure.

This leads to the famous tech industry buzzword: "The model you're using today is its dumbest day ever."

Tara cut to the chase, pointing out that the most absurd part of this statement is that while everyone verbally agrees, when it comes to actually getting work done, they still unconsciously schedule based on the level of this "most basic" model.

To break this deadly human inertia, OpenAI internally slashed its rigid annual roadmap, forcibly shortening its planning horizon to two or three months.

Tara said in the interview that products must be built according to a model that looks two to three months ahead—betting on the present or gambling on a year from now will both fail.

But these past two or three months were not about blind trial and error based on guesswork—internal processes strictly prohibit any “black boxes.”

Tara's real strength is: avoiding long-term nominal planning, but ensuring the product strictly follows the research team's actual progress.

The product and research teams are required to maintain extremely frequent communication to fully align on one key question: In which specific capabilities will the model improve next, and exactly when will these improvements be delivered?

After understanding this underlying detail, the product can agilely iterate in sync with technological breakthroughs.

This isn’t abandoning the roadmap—it’s turning that rigid plan on the wall into a real-time navigation system that adapts continuously based on frontline conditions.

Tara previously came from Stripe, a company that highly values writing culture, and was accustomed to refining her briefs to perfection before sending them out.

But at OpenAI, she completely changed her habit: writing long documents only for herself.

Because "complete materials" no longer signify "deep thinking," the intellectual credibility of lengthy documents has become zero.

Now, working prototypes, A/B test data, and user feedback have completely replaced overly lengthy, inflated documentation.

But she maintains a highly effective work principle: when a plan is 70% complete, take it to the person who needs to authorize it, and let them help you finish the remaining 30%.

The reason is practical: when a flawless, perfect plan is laid out, others' ideas will simply bounce off it; only by keeping some rough edges will others be willing to reach out and polish it with you.

Tara divides writing into two categories:

• Report-style writing (weekly reports, summaries, abstracts, format conversions): Hand it all over to AI—automate whenever possible;
• Thoughtful writing (project rationale, route selection, dispute resolution): Never outsourced to AI.

Why can't you give thinking-based writing to AI?

The entire process of outlining, refining your writing, and revising repeatedly is itself the process of clarifying your thoughts.

Assigned to AI writing—the words are there, but the step of “thinking it through” never actually happened in your mind.

To prevent cognitive decline, she imposed two strict rules on herself:

You must read a document as many times as the number of people who will read it.

2. The amount of time needed for preparation must add up to the total time the meeting will take.

When AI has lowered the barriers to design, analysis, coding, and prototyping to an extremely low level, tasks that once required a team can now be roughly completed by a single person.

What bridges the gap between people?

"Ambition".

True advanced players don't just use AI to automate repetitive tasks—they use it to infinitely expand the boundaries of their own capabilities.

This also completely transforms the responsibilities of product managers (and all knowledge workers):

In the past: Effort was spent on scheduling, writing documentation, and reviews—peripheral rituals.

Now: Identify the critical issues that determine success or failure, make the most pointed assumptions, and validate them at lightning speed.

Most importantly, raise the team's ambition.

When someone proposes a solution, your value lies in asking: “Can the ceiling be even higher? Can the speed be increased tenfold?”

Three insider slang terms at OpenAI

These three sentences correspond exactly to three self-checks.

Can this be done any faster?

Ask yourself if your ambition is enough and whether your goals can be raised even higher.

The third point is the toughest: Are you using this product every single day since you woke up? Can you no longer live without it? Have you bet all your taste and judgment on whether it’s truly good to use?

Lenny asked her what internally changed over the past three to six months that caused the trend on Twitter to clearly shift from Claude Code to Codex.

Tara's response was: Nothing has been changed. The team has been using it intensively internally, continuously gathering feedback and iterating quickly—the way we work has remained unchanged from start to finish.

The only change is that people outside have started to notice.

Coding is results-driven: After the Agent modifies the code, it runs the tests—either it passes, or it doesn’t.

But knowledge work is completely different.

A final PPT states a "90% success rate"—you can't simply believe it based on the number alone.

You must examine what its original input was and how the reasoning progressed step by step.

This is why the core design of ChatGPT Work isn't to deliver a polished final product outright, but rather to lay out all references, inputs, and partially completed tasks so you maintain full control throughout the process.

Only 20 minutes remain until the client meeting; ChatGPT Work delivered page 8 of the presentation in 2 minutes and 30 seconds, with all referenced materials fully visible.

Everyone has access to the tools; foundational capabilities are being rapidly equalized.

In the age when AI takes over rowing, the only thing that truly determines your value is whether you have the skill to steer.

Where do you want to take this ship?

Reference materials:

https://x.com/lennysan/status/2094444052204998908

https://youtu.be/zMvBMfj4cSQ

This article is from the WeChat public account "New Intelligence Yuan," authored by ASI Revelation, edited by Yuan Yu.