diff --git a/docs/random/streamheader b/docs/random/streamheader
new file mode 100644
index 0000000000..31bff21510
--- /dev/null
+++ b/docs/random/streamheader
@@ -0,0 +1,83 @@
+A lot of data streams need a few initial buffers to make sense of the
+data stream.  With these initial buffers, it can pick up at any point in
+the rest of the data stream.
+
+Exampes:
+- Vorbis and Theora have three initial Ogg packets
+- MPEG4 has codec data
+- GDP protocol serializes the initial new_segment event and the initial caps
+
+It is important that elements that get connected after the stream starts,
+receive these initial buffers.  Also, sink elements should know that these
+initial buffers are necessary for every new "client" of the sink element;
+for example, multifdsink needs to send these initial buffers before the
+normal buffers.
+
+Currently, this is done by putting a 'streamheader' value on the caps.
+'streamheader' is a GST_TYPE_ARRAY of GstBuffer, ie. an ordered list of
+buffers.
+
+The buffers themselves still get pushed as normal buffers, but with the
+IN_CAPS flag set.  This allows elements that do not know about streamheader
+to still function correctly (ie. filesink and others)
+
+It is of course important that these streamheader buffers are not sent twice
+(once because they're in the caps, and once because they're received as
+buffers).
+
+So, an element that is aware of streamheader in the caps should probably best
+drop all incoming IN_CAPS buffers.
+
+Rules for sending streamheaders:
+- all streamheader buffers should have IN_CAPS set
+- from this list of streamheader buffers, a "streamheader" caps field should
+  be created as a GST_TYPE_ARRAY of GstBuffer.  This should be done by
+  copying the buffers.
+  (the only important thing about the buffers in this list is the data;
+   other attributes of the buffers get ignored)
+- each of these buffers should then have these caps set on them.
+- when all streamheader buffers are collected in the element, pad caps should
+  be set, including this streamheader
+- streamheader buffers should be sent consecutively, and before any of the
+  data (non-IN_CAPS) buffers they apply to.
+- when new streamheader buffers need to be pushed out, this process is
+  repeated.  Receiving a new IN_CAPS buffer after a non-IN_CAPS buffer
+  signifies resetting streamheader, as does the new set_caps with different
+  streamheader right before.
+
+Elements that can send streamheader caps:
+- vorbisenc
+- theoraenc
+- gdppay
+
+Elements that can receive streamheader caps:
+- multifdsink
+
+Elements that can receive and send streamheader caps:
+- oggmux
+
+Elements that could/should use this:
+- ffenc_mpeg4 (currently uses/sets codec_data)
+- theoradec, vorbisdec (currently only look at incoming buffers)
+- oggdemux
+
+FUTURE PLANS
+------------
+The concept of streamheader is more generally applicable.  We may want to
+find a better way of implementing this than having two ways of telling
+downstream elements about them.  Currently an element that is
+streamheader-aware needs to look both at caps and incoming buffers.
+
+One option would be to only rely on the buffer flow, and make all elements
+aware of a HEADER flag, causing them to keep these buffers around and push
+them for each new pad link.  This needs to be done in every element; a
+scenario like
+  videotestsrc ! theoraenc
+where later on an oggmux gets connected, should still work without dropping
+these header buffers.  But similarly,
+  videotestsrc ! theoraenc ! identity
+with a later connection of oggmux to identity should work.
+
+This could tie in with an idea to have pads store some initial data (the
+first new_segment event, tag events, ...), which at the moment also get
+dropped on dynamic in-stream links.