diff --git a/QUERY-SPEC.md b/QUERY-SPEC.md
new file mode 100644
index 0000000..bc7f857
--- /dev/null
+++ b/QUERY-SPEC.md
@@ -0,0 +1,131 @@
+# KDL Query Language Spec
+
+The KDL Query Language is a small language specially tailored for querying KDL
+documents to extract nodes and even specific data. It is loosely based on CSS
+selectors for familiarity and ease of use. Think of it as CSS Selectors or
+XPath, but for KDL!
+
+## Selectors
+
+Selectors use selection operators to filter nodes that will be returned by an
+API using KQL. The main differences between this and CSS selectors are the
+lack of `*` (use `[]` instead), and the specific syntax for
+[matchers](#matchers) (the stuff between `[` and `]`), which is similar, but not identical to CSS.
+
+* `a > b`: Selects any `b` element that is a direct child of an `a` element.
+* `a b`: Selects any `b` element that is a _descendant_ of an `a` element.
+* `a b || a c`: Selects all `b` and `c` elements that are descendants of an `a` element. Any selector may be on either side of the `||`. Multiple `||` are supported.
+* `a + b`: Selects any `b` element that is placed immediately after a sibling `a` element.
+* `a ~ b`: Selects any `b` element that follows an `a` element as a sibling, either immediately or later.
+* `[accessor()]`: Selects any element, filtered by [an accessor](#accessors). (`accessor()` is a placeholder, not an actual accessor)
+* `a[accessor()]`: Selects any `a` element, filtered by an accessor.
+* `[]`: Selects any element (a )
+
+## Matchers
+
+Matchers are used to filter nodes by their various attributes (such as values,
+properties, node names, etc). With the exception of `top()`, they are all used
+inside a `[ ]` selector. Some matchers are unary, but most of them involve
+binary operators.
+
+* `top()`: Returns all toplevel children of the current document.
+* `top() > []`: Equivalent to `top()` on its own.
+* `[val()]`: Selects any element with a value.
+* `[val(1)]`: Selects any element with a second value.
+* `[prop(foo)]`: Selects any element with a property named `foo`.
+* `[prop]`: Selects any element with a property named `prop`.
+
+Attribute matchers support certain binary operators:
+
+* `[val() = 1]`: Selects any element whose first value is 1.
+* `[prop(name) = 1]`: Selects any element with a property `name` whose value is 1.
+* `[name = 1]`: Equivalent to the above.
+* `[name() = "hi"]`: Selects any element whose _node name_ is "hi". Equivalent to just `hi`, but more useful when using string operators.
+* `[val() != 1]`: Selects any element whose first value exists, and is not 1.
+
+The following operators work with any `val()`, `prop()`, or `name()` values.
+If the value is not of the same type, the operator will always fail ("1" is
+never coerced to 1, and there is no "universal" ordering across all types.):
+
+* `[val() > 1]`: Selects any element whose first value is greater than 1.
+* `[val() >= 1]`: Selects any element whose first value is greater than or equal to 1.
+* `[val() < 1]`: Selects any element whose first value is less than 1.
+* `[val() <= 1]`: Selects any element whose first value is less than or equal to 1.
+
+The following operators work only with string `val()`, `prop()`, or `name()` values. If the value is not a string, the matcher will always fail:
+
+* `[val() ^= "foo"]`: Selects any element whose first value starts with "foo".
+* `[val() $= "foo"]`: Selects any element whose first value ends with "foo".
+* `[val() *= "foo"]`: Selects any element whose first value contains "foo".
+
+## Map Operator
+
+KQL implementations MAY support a "map operator", `=>`, that allows selection
+of specific parts of the selected notes, essentially "mapping" over a
+selector's result set.
+
+Only a single map operator may be used, and it must be the last element in a
+selector string.
+
+The map operator's right hand side is either an [`accessor`](#accessors) on
+its own, or a tuple of accessors, denoted by a comma-separated list wrapped in
+`()` (for example, `(a, b, c)`).
+
+## Accessors
+
+Accessors access/extract specific parts of a node. They are used with the [map
+operator](#map-operator), and have syntactic overlap with some
+[matchers](#matchers).
+
+* `name()`: Returns the name of the node itself.
+* `val(2)`: Returns the third value in a node.
+* `val()`: Equivalent to `val(0)`.
+* `prop(foo)`: Returns the value of the property `foo` in the node.
+* `foo`: Equivalent to `prop(foo)`.
+* `props()`: Returns all properties of the node as an object.
+* `values()`: Returns all values of the node as an array.
+
+## Examples
+
+Given this document:
+
+```kdl
+package {
+    name "foo"
+    version "1.0.0"
+    dependencies platform="windows" {
+        winapi "1.0.0" path="./crates/my-winapi-fork"
+    }
+    dependencies {
+        miette "2.0.0" dev=true
+    }
+}
+```
+
+Then the following queries are valid:
+
+* `package name`
+    * -> fetches the `name` node itself
+* `top() > package name`
+    * -> fetches the `name` node, guaranteeing that `package` is in the document root.
+* `dependencies`
+    * -> deep-fetches both `dependencies` nodes
+* `dependencies[platform]`
+    * -> fetches any dependencies nodes with a `platform` prop (just the one, in this case)
+* `dependencies[prop(platform)]`
+    * -> Identical to the above. Plain identifiers are equivalent to `prop(<identifier>)`.
+* `dependencies > []` ->
+    * fetches all direct-child nodes of any `dependencies` nodes in the
+      document. In this case, it will fetch both `miette` and `winapi` nodes.
+
+If using an API that supports the [map operator](#map-operator), the following
+are valid queries:
+
+* `package name => val()`
+    * -> `["foo"]`.
+* `dependencies[platform] => platform`
+    * -> `["windows"]`
+* `dependencies > [] => (name(), val(), path)`
+    * -> `[("winapi", "1.0.0", "./crates/my-winapi-fork"), ("miette", "2.0.0", None)]`
+* `dependencies > [] => (name(), values(), props())`
+    * -> `[("winapi", ["1.0.0"], {"platform": "windows"}), ("miette", ["2.0.0"], {"dev": true})]`