diff --git a/docs/src/api.md b/docs/src/api.md
index 378bf72a..434ee70a 100644
--- a/docs/src/api.md
+++ b/docs/src/api.md
@@ -85,4 +85,5 @@ It is defined in Functors.jl and re-exported by Optimisers.jl here for convenien
 Functors.KeyPath
 Functors.haskeypath
 Functors.getkeypath
+Functors.setkeypath!
 ```
diff --git a/src/interface.jl b/src/interface.jl
index ac9b90bc..be0427ab 100644
--- a/src/interface.jl
+++ b/src/interface.jl
@@ -67,6 +67,12 @@ function update(tree, model, grad, higher...)
   update!(t′, x′, grad, higher...)
 end
 
+function update!(::AbstractRule, model, grad, higher...)
+  throw(ArgumentError("""update! must be called with an optimiser state tree, not a rule.
+  Call `opt_state = setup(rule, model)` first, then `update!(opt_state, model, grad)`.
+  """))
+end
+
 function update!(tree, model, grad, higher...)
   # First walk is to accumulate the gradient. This recursion visits every copy of
   # shared leaves, but stops when branches are absent from the gradient:
diff --git a/test/runtests.jl b/test/runtests.jl
index fc0fe57f..bb5cbec9 100644
--- a/test/runtests.jl
+++ b/test/runtests.jl
@@ -104,6 +104,10 @@ end
       @test isnan(m3n.γ[3])
     end
 
+    @testset "friendly error when using rule instead of state" begin
+      @test_throws ErrorException Optimisers.update!(Adam(), rand(2), rand(2))
+    end
+
     @testset "Dict support" begin
       @testset "simple dict" begin
         d = Dict(:a => [1.0,2.0], :b => [3.0,4.0], :c => 1)