fill out features section

Author: WardBrian

paper/paper.bib

@@ -87,3 +87,19 @@ @inproceedings{emscripten2011
     location = {Portland, Oregon, USA},
     series = {OOPSLA '11}
 }
+
+
+@article{Kumar_ArviZ_a_unified,
+    author = {Kumar, Ravin and Carroll, Colin and Hartikainen, Ari and Osvaldo, Martin},
+    doi = {10.21105/joss.01143},
+    journal = {Journal of Open Source Software},
+    title = {{ArviZ a unified library for exploratory analysis of Bayesian models in Python}}
+}
+
+@Misc{posterior2025,
+    title = {{posterior}: Tools for Working with Posterior Distributions},
+    author = {Paul-Christian Bürkner and Jonah Gabry and Matthew Kay and Aki Vehtari},
+    year = {2025},
+    note = {R package version 1.6.1},
+    url = {https://mc-stan.org/posterior/},
+}

paper/paper.md

@@ -49,9 +49,8 @@ working C++17 development environment installed in order to use Stan.
 This requirement can be difficult to satisfy for users from less technical backgrounds or with
 IT-imposed constraints on their systems. Installation-related questions are
 ubiquitous on the Stan forums. This can be a particular pain point in learning environments,
-where instructors do not
-want to spend a substantial fraction of their teaching time ensuring students have a working
-environment and copy of the relevant data and models.
+where instructors do not want to spend a substantial fraction of their teaching time just to ensure
+students have a working environment and copy of the relevant data and models.
 
 Thanks to its browser-based design requiring no installation, Stan Playground allows users
 to immediately focus on the issues of modeling and analysis that matter to them.
@@ -63,7 +62,7 @@ immediately, regardless of their local system configuration.
 A similar level of convenience could be achieved by performing both compilation and sampling on
 centrally provided server, but there are several issues with this approach.
 Sampling involves executing user-written code that could perform potentially-unbounded amounts of
-compute. Offering this as a public server would require authentication or rate-limiting.
+compute. Offering this as a public service would require authentication or rate-limiting.
 Additionally, even in a restricted language like Stan, one would have to take serious precautions before
 allowing untrusted user code to run unchecked.
 
@@ -77,30 +76,48 @@ technical, process into loading a web page in any modern browser.
 
 # Key features
 
-<!-- screenshot -->
-
 ![Stan Playground's UI after running the Lotka-Volterra example\label{fig:screenshot}](screenshot.png)
 
+Stan Playground provides many features to allow end-to-end analysis projects to be completed in
+the browser.
+
+1. **Data Preparation**. The input data for the Stan model can be provided manually as JSON or
+   generated by Python or R scripts which can download existing datasets or dynamically generate fake data.
+
+2. **Model Editing**. The Stan editor in Stan Playground features integration with the Stan compiler
+   to provide on-the-fly diagnostics and formatting.
 
-- data generation
-- model editing (with diagnostics etc), compilation, sampling
-- downstream analysis
-- sharing features
-- user compilation server
+3. **Compilation**. A premier feature of Stan Playground is the ability to compile and run without
+   needing additional tools installed. This is achieved through a server that handles this compilation
+   and provides WebAssembly modules back to the client. While a public server is hosted by the Flatiron Institute,
+   users can also run their own and connect to that instead in the settings UI.
 
+4. **Sampling**. User can control the key parameters for the sampling algorithm (number of iterations, etc) and
+   monitor progress in real time.
+
+5. **Posterior Analysis**. Stan Playground provides many built-in diagnostics and visualizations for the
+   posterior sample. Additionally, users can write Python or R scripts to leverage the existing ecosystem of
+   tools such as ArviZ [@Kumar_ArviZ_a_unified] or `posterior` [@posterior2025].
+
+6. **Sharing and Export**. Stan Playground by default saves your analysis in browser local storage.
+   For longer-term storage, users can export their projects to zip files for local use, or upload
+   to a Github Gist. Github Gist links facilitate easy sharing, as they can be used in URL parameters
+   to load a specific project when a link is clicked. More customizable sharing options are also available
+   through URL parameters.
 
 # Examples
 
 Stan Playground provides several example programs on the left
 sidebar of the home page at various levels of complexity, spanning from a basic
 linear regression on fake data to
 [a reimplementation of the model and analysis](https://stan-playground.flatironinstitute.org/?project=https://gist.github.com/WardBrian/d8ab811b137085f154b6145d3c36cbc4)
-from @Carpenter2018.
+from @Carpenter2018. The above link uses Stan Playground's sharing features to pre-populate the project on page load.
 
 # Implementation details
 
 Stan Playground consists of two components. The first is a server that can compile Stan
 models to WebAssembly, implemented in Python and leveraging the existing Emscripten toolchain [@emscripten2011].
+This server is also published as a [Docker](https://www.docker.com/) image if users wish to run their own local copy.
 
 The remainder of the project is the user-facing web application, which can be found at
 [https://stan-playground.flatironinstitute.org/](https://stan-playground.flatironinstitute.org/).
@@ -110,7 +127,6 @@ panels to allow editing of Stan models, compilation, sampling, and downstream an
 results. The use of WebAssembly and web workers allows these computations to complete quickly
 and without freezing the user interface.
 
-
 # Acknowledgements
 
 Andrew Gelman and Jonah Gabry (both: Columbia University)