Working with data is a collaborative discipline. Fast exploration requires input from multiple stakeholders and domain experts. Working with others should be a delight, not a burden.

Collaboration needs to happen inside the medium, not on external collaboration platforms. Data notebooks should be collaborative by nature, with built-in tools for commenting, discussing, reviewing, and merging changes made by multiple users.

Rather than relying on external mechanisms such as git, the notebook should implement versioning natively.

Collaboration should be supported within the data team, but also outside of the data team, available to non-technical users and citizen data scientists.

Data notebooks should be used to support decision-making, with clear links between data and the conclusions drawn from it.

Rather than following a fairly straightforward project lifecycle from prototype to production, as known from the engineering world, notebooks need to allow free exploration, scenario analysis, and frequent branching into different directions.

In notebooks, exploration is seen as a first-class citizen. Exploration is a different type of thinking and a data notebook needs to facilitate the exploratory mindset.

Unlike traditional code editors, which are optimized for time-to-production, notebooks are optimized for time-to-insight. They decrease the feedback loop and allow fast iterations, immediately recomputing insights based on the user input. Data notebooks should be interactive, allowing users to explore and analyze data in a dynamic and engaging way.

Notebook should not exclude its users based on their technical abilities. Instead, they should be progressive — easy to start with, letting users discover more advanced features only when needed. Code as the most technical element of the notebook should be optional. Both technical and non-technical audiences should be able to work the notebook, in any way they choose. Notebook should assist users with tasks they couldn’t do before, and improve their productivity with tasks they could. Notebooks should not stand in the way.

Keeping the barrier low means more people will be able to get the insights they need. It democratizes the access to data. Instead of having to rely on specialized roles to extract needed insights, users can self-serve and get the answers directly themselves.

While the barrier is low, the notebook should not place an artificial limit on what’s possible to achieve with it. When needed, the user should be able to work within the notebook to develop apps and models of any complexity.

Where spreadsheets start to run into their limitations, notebooks should be able to continue working and scaling with the complexity of the task. They should be as powerful as any IDE — having no limit on what you can build.

All data work should be easily reproducible by others.

Notebooks should act as a self-sufficient artifact where code, data, and documentation are stored in one place. Every notebook should be runnable by all the collaborators, regardless of their choice of operating system or hardware.

Notebooks should be easy to pick up by collaborators and extended. They should encourage easy last-mile analysis, allowing even non-technical users to continue working with the work of others and explore the effect of different parameters.

Rather than creating a separate interface for data analysts (e.g. a code editor) and a separate interface for the consumers of those insights (e.g. dashboards, slides), a perfect tool for data exploration should act as a universal medium for both production and consumption of insights.

This is similar to spreadsheets, where the environment used to create a model is the same as the environment used to explore the model.

Unifying the product and consumption environment leads to a tighter feedback loop and allows consumers to easily extend prior work and explore different scenarios on their own, without having to rely on the creator to make the desired changes.

The interface should allow for the creation of polished artifacts that can be used in production either as a replacement for a traditional manually-built app or interacting with it programmatically as an API to compose multiple notebooks together into a more complex workflow.

Since the quality of context drives the quality of results, the medium needs to be able to encompass all kinds of relevant assets, both in structured and unstructured form, in an organized manner.

Data analysis is not about code, but about insights. To find an insight, we need both computation and storytelling. Notebooks allow for both to happen in the same medium.

Every analysis should be transparent, with clear explanations of methods and assumptions used in the analysis.

Notebooks should allow to be used as a knowledge base.

Insights are meant to be shared and distributed. The role of a notebook is to make sharing easier. They need to be discoverable.

Notebooks should be easy to search through and easy to organize. They should prevent double work and allow users to discover insights already produced by other team members.

Exploration is a messy process, prone to leaking data. Traditionally, data is scattered across many places with poor access control such as Downloads folder, email attachments, or one-off S3 buckets.

Data notebooks should be secure by default, with appropriate controls in place to protect sensitive data.

Given the universal nature of the notebook, there is no need to use external tools to work with data. Everything happens within the notebook environment where every action is audited and the data doesn’t leak outside of a secure perimeter.

While notebooks shine in exploratory programming and exploratory data analysis, they should retain their universal nature allowing their users to build any type of project from analysis to large machine learning projects.

This interface should be easily extensible and compatible with any kind of asset — text, executable code, visualizations, maps, or embedded files.

Notebooks should support both traditional but also AI-assisted computing, and be easily adaptable to novel advances in AI computing, where humans and AI agents can work side-by-side.