diff --git a/doc/http-parsing.txt b/doc/http-parsing.txt
index 21c3fc5..7e731c8 100644
--- a/doc/http-parsing.txt
+++ b/doc/http-parsing.txt
@@ -212,3 +212,28 @@
 recognizable by the recipient. Unrecognized header fields SHOULD be ignored by
 the recipient and MUST be forwarded by transparent proxies.
 
+---- The correct way to do it ----
+
+- one http_session
+  It is basically any transport session on which we talk HTTP. It may be TCP,
+  SSL over TCP, etc... It knows a way to talk to the client, either the socket
+  file descriptor or a direct access to the client-side buffer. It should hold
+  information about the last accessed server so that we can guarantee that the
+  same server can be used during a whole session if needed. A first version
+  without optimal support for HTTP pipelining will have the client buffers tied
+  to the http_session. It may be possible that it is not sufficient for full
+  pipelining, but this will need further study. The link from the buffers to
+  the backend should be managed by the http_request, provided that they are
+  serialized. Each http_session, has 0 to N http_requests. Each http_request
+  belongs to one and only one http_session.
+
+- each http_request has 1 request message, and 0 or 1 response message. Each of
+  them has 1 and only one http_request. An http_request holds informations such
+  as the HTTP method, the URI, the HTTP version, the transfer-encoding, the
+  HTTP status, the authorization, the req and resp content-length, the timers,
+  logs, etc... The backend and server which process the request are also known
+  from the http_request.
+
+- both request and response messages hold header and parsing informations, such
+  as the parsing state, start of headers, start of request, captures, etc...
+