Research Workflow with OpenAI Assistants

Nikita Melashchenko · (27 November 2024) · guide

For most of 2023 and into early 2024, working with ChatGPT meant opening a browser tab, navigating to chat.openai.com, and typing or pasting text into a chat window. The workflow for a researcher or an academic was straightforward but clunky: copy a paragraph from your draft, switch to the browser, paste it into the chat, read the response, switch back to your document, reflect and continue writing. The AI was a separate destination, somewhere you went to, not something that was part of your workspace.

In 2024, OpenAI changed this interaction paradigm twice. First, in May, by releasing a desktop application that brought AI out of the browser and onto the desktop. Then, in November, by introducing a feature called Work with Apps that connected ChatGPT directly to the applications where work actually happens such as code editors, text editors and note-taking apps. Together, these two developments represent a shift in how information workers, and academics in particular, can integrate AI into their daily practice.

The desktop app: AI leaves the browser

In May 2024, OpenAI released the ChatGPT desktop app for macOS. At first glance, it looked like a minor convenience. The same chat interface, but in a standalone window instead of a browser tab. The real innovation was a keyboard shortcut — Option + Space. With that combination, ChatGPT could be summoned from anywhere on the system, regardless of what application was in the foreground. It was a system-level utility, always one keystroke away.

For researchers and writers, this was a meaningful change. The friction of context-switching, leaving your document, opening a browser, finding the right tab, disappeared. Asking the AI a quick question while writing became as natural as opening Spotlight to search for a file. The cognitive cost of consulting the AI dropped significantly, which meant it happened more often and more naturally.

The shift was subtle but important. From “going to the AI” to “the AI coming to you.”

Work with apps: AI enters the workflow

In November 2024, OpenAI took this a step further with a feature called Work with Apps. The ChatGPT desktop app could now read content from a handful of applications — VS Code, Xcode, TextEdit and Terminal. When you opened the ChatGPT window with one of these apps in the foreground, the assistant could see what was on your screen. It could read the code you were editing, the text you were drafting, or the terminal output you were troubleshooting.

By December 2024, the list of supported applications expanded to include PyCharm, IntelliJ IDEA, Apple Notes and Notion.

The practical implication was no more copy-pasting. A researcher writing a paper in VS Code — using Markdown or Quarto — could summon ChatGPT, and the assistant would already have the context of the document open in the editor. You could ask “Does this paragraph flow well?” or “Can you suggest a transition to the next section?” without having to select, copy, switch windows, paste and explain what the text was about. The AI could see it.

This is what made the feature transformative rather than merely convenient. The interaction was no longer about describing your work to the AI; it was about the AI observing your work and responding to it directly.

What this means for text professionals and academics

The traditional users of code editors and IDEs were programmers. But over the past several years, tools like VS Code have become workspaces for a much broader group of information workers. With extensions for Markdown, Quarto and LaTeX, a code editor is now a powerful environment for academic writing, complete with version control, integrated previews and extensible tooling. The arrival of AI integration within these environments is significant precisely because it meets these users where they already work.

Writing and editing. For an academic drafting a journal article in VS Code, the AI can now review a paragraph in real time without requiring the writer to leave the editor. You can ask for feedback on clarity, suggest alternative phrasing, or check whether an argument follows logically from the preceding section — all while the document remains open and in focus.

Research scripting. Academics who write data analysis scripts in Python or R can get inline help without context-switching. The AI sees the code in the editor and can suggest corrections, explain error messages, or help refactor a function — directly in the context of the project, not in a disconnected chat window where the code has to be explained from scratch.

Note-taking and synthesis. The integration with Apple Notes and Notion means that research notes, such as reading summaries, conference observations, annotated bibliographies, are directly accessible to the AI. You can ask the assistant to synthesise themes across your notes or identify gaps in your reading without having to manually compile and paste the relevant material.

The IDE as a universal workspace. The broader point is that the boundary between “developer tool” and “writing tool” has been dissolving for some time. What the AI integration does is accelerate this. A tool that was once the exclusive domain of programmers is now a workspace where any information worker can write, edit, analyse and collaborate with an AI assistant — all in the same environment.

The interface evolution

It is worth stepping back to consider the trajectory of AI interfaces over the past two years:

This trajectory mirrors a pattern familiar from the history of computing itself. Computing moved from mainframes (go to the computer) to personal computers (the computer on your desk) to smartphones (the computer in your pocket). AI appears to be following the same trajectory, from a centralised service you visit, to a local application you summon, to an embedded capability woven into your existing tools, compressed into roughly two years.

Where to next

The trajectory described above — from browser to desktop to embedded — is unlikely to stop at reading the contents of a few supported applications. The direction points toward deeper integration: AI that can not only observe what is on the screen but interact with the application directly, editing text, running commands, navigating between files. Early signs of this are already visible in tools like GitHub Copilot, which has been embedded in VS Code since 2022, and Cursor, an AI-native code editor that gained significant traction in 2024. In both cases, the AI operates within the editor as a participant rather than an observer.

For academia, this raises a question that the sector has been grappling with since the release of ChatGPT in late 2022: where does legitimate assistance end and academic misconduct begin? The concern is not new. Spell checkers, grammar tools, statistical software and even search engines have all prompted similar debates at the time of their introduction. But AI is different in degree, if not in kind, because it can generate text, construct arguments and produce analysis that is difficult to distinguish from human-authored work.

The stigma around AI use in academic settings is real and, in many cases, well-founded. Universities and journals have responded with policies ranging from outright bans to disclosure requirements, and the inconsistency across institutions reflects genuine uncertainty about how to draw the line. A researcher who uses AI to check grammar is doing something many would consider acceptable. A researcher who uses AI-generated sections is doing something many would consider problematic. But between those extremes lies a wide grey area, using AI to restructure an argument, to suggest alternative framings, to identify logical gaps, where reasonable people disagree. Perhaps, it is the extent and quality of human oversight that should be the measure of acceptability, rather than the mere presence of AI assistance.

What the interface evolution described in this post does is make these questions more pressing, not less. When AI was a separate browser tab, the act of consulting it was deliberate and bounded. When AI is embedded in the writing environment, always present and already aware of the document, the boundary between the author’s work and the AI’s contribution becomes harder to delineate, both for external observers and for the author themselves. This problem will likely become more acute as operating systems and applications embed AI more deeply, and as the AI’s ability to observe and interact with the user’s work becomes more sophisticated.

None of this will slow adoption. As AI becomes embedded at the operating system level, woven into the tools people already use every day, the distinction between “using AI” and “using a computer” will continue to blur. The stigma that currently surrounds AI use in academic settings is real, but it exists in tension with a technological trajectory that is making AI assistance an ambient feature of the working environment rather than a deliberate choice to invoke a separate tool. The question of how academia reconciles these two forces remains open, but the direction of the technology itself is not.

Closing thoughts

The shift from browser to desktop to embedded AI is not just a convenience improvement. It changes the nature of human-AI collaboration. When the AI can see what you are working on, the interaction becomes a conversation about the work itself rather than a conversation about explaining the work to the AI first. For researchers, writers and academics, this removes what was perhaps the most significant barrier to productive AI use, the friction of context.

The tools are still evolving, and the boundaries of what the AI can and cannot observe will continue to expand. But the direction is clear. AI is moving from being a destination to being an environment, a layer that sits within the tools we already use, available when needed, aware of what we are doing and ready to help.

Also see

Using the OpenAI Assistant as a Telegram Bot

Using the OpenAI Assistant in a Jupyter Notebook with PyCharm