diff --git a/dev_docs/design_principles.md b/dev_docs/design_principles.md
new file mode 100644
index 00000000..695d72ef
--- /dev/null
+++ b/dev_docs/design_principles.md
@@ -0,0 +1,17 @@
+# Design Principles (HCI)
+
+cellxgene is a tool for scientists investigating single cell rna seq datasets. The design of cellxgene proceeds from a small number of core principles:
+
+0. the tool, first and foremost, must produce views that are **scientifically valid** at all times
+1. the tool should be highly **scaleable** and handle *exploration* of millions of cells in the browser at interactive speeds, including interactive crossfiltering and dataframe subsetting
+2. the tool should be **data dense** and give scientists powerful views into data
+3. the tool should be powerfully expressive and **optimized for the `nth` day of use** rather than the first day, in the spirit of enterprise tools, even if that requires training or onboarding 
+    - (think: photoshop, which is supported by a broad array of youtube tutorials, books and trainings)
+4. the tool should **avoid duplicating data** onscreen 
+    - for example, if we render a categorical label on the left hand side bar, locate further information related to that label in place on the left hand sidebar, rather than rendering that data again.
+    - for a concrete example, see the relationship between this solution: https://github.com/chanzuckerberg/cellxgene/pull/827 and this problem: https://github.com/chanzuckerberg/cellxgene/issues/762
+5. the tool should enable users to rapidly test hypotheses on the application, which may require it to **enable interactive recomputation of views** into the data, for example: 
+    - interactive differential expression
+    - interactive reprojection of umap
+6. the tool should enable a productive workflow between those who are computational and those who are not
+7. the tool should **minimize extraneous use of color**, leaving color to primary workflow actions like `compute differential expression` or `create new categorical metadata` or `create new geneset`. This leaves the color space to `colorby` actions, such as `colorby geneset` and `colorby categorical field`