diff --git a/proposals/4139-bot-buttons.md b/proposals/4139-bot-buttons.md
new file mode 100644
index 00000000000..5ee6e758517
--- /dev/null
+++ b/proposals/4139-bot-buttons.md
@@ -0,0 +1,201 @@
+# MSC4139: Bot buttons & conversations
+
+Nearly all bots and bridges in the Matrix ecosystem use a text-based interface to support their
+operations. These interfaces are typically highly structured commands and require the user to know
+the entire incantation for the action they want to invoke, making them feel like "power user"
+features.
+
+Further, interacting with bots today is extremely transactional: the user sends a command and the
+bot performs the action as-is or spews errors back at the user due to a typo. If an error was
+returned, the entire command needs to be re-run.
+
+A more user-friendly approach is to have the user provide the bot with information as needed,
+without having to guess at the bot's current state. This proposal calls such an approach a
+"conversation" with the bot - the user does something to "start" the conversation, and the bot
+provides a limited set of prompts to continue the conversation. This repeats until the conversation
+ends (usually by the bot saying so explicitly). Users may hold multiple concurrent conversations
+with bots. Conversation starters are deliberately left as a bot implementation detail in this
+proposal to allow the ecosystem to explore this new interaction technique. Examples may include the
+user opening a DM with the bot, sending a `!command` message, or, in future, sending a slash command
+like `/start`.
+
+This conversation approach is heavily inspired by platforms like Telegram.
+
+## Proposal
+
+A new `m.prompts` [mixin](https://github.com/matrix-org/matrix-spec-proposals/blob/main/proposals/1767-extensible-events.md#mixins-specifically-allowed)
+is specified which describes actions another user in the room can take to further the conversation.
+
+The `m.prompts` mixin contains some scoping parameters, rendering hints, and the actual prompts
+themselves. For example, when applied to an `m.message` event, the `m.prompts` may look like the
+following:
+
+*Note*: The JSON comments are normative, and irrelevant fields are not shown.
+
+```jsonc
+{
+  "type": "m.message",
+  "sender": "@bot:example.org",
+  "content": {
+    "m.text": [
+      {"body": "Hello! Say <code>!roll [dice]</code> to roll some dice.", "mimetype": "text/html"},
+      {"body": "Hello! Say `!roll [dice]` to get started."}
+    ],
+    "m.prompts": {
+      // Clients which recognize `m.prompts` would use `intro` to render the event instead. This
+      // allows the remainder of the event to be a fallback for unsupported clients.
+      "intro": {
+        "type": "m.message",
+        "content": {
+          "m.text": [
+            {"body": "Hello! What would you like to roll today?"}
+          ]
+        }
+      },
+      // These are the users who should see the `prompts`. Other users may see something like "you
+      // do not have permission to reply to this message" instead of prompts. `scope` is optional:
+      // when not supplied, all users who can see the message can respond. When an empty array, no
+      // one can respond. Clients SHOULD NOT show prompts to users who are descoped.
+      "scope": [
+        "@alice:example.org",
+        "@bob:example.org",
+      ],
+      // These are the options a user has. Note the 2 distinct types and 3 label approaches.
+      "prompts": [
+        {
+          // `type` is the prompt type: "preset" (show a button) or "input" (shown below)
+          "type": "preset",
+          // `id` is used by the bot to figure out what prompt the user picked. It is an opaque ID.
+          "id": "1d6",
+          // `label` is an extensible event with deliberately no `type`.
+          "label": {
+            "m.text": [{"body": "1 six sided die"}]
+          }
+        }, {
+          "type": "preset",
+          "id": "surprise",
+          "label": {
+            // This should render as an image event, hopefully
+            // Requires https://github.com/matrix-org/matrix-spec-proposals/pull/3552
+            "m.text": [{"body": "🎲❓"}], // fallback
+            "m.file": {
+              "url": "mxc://example.org/abc123"
+            },
+            "m.image_details": {
+              // Clients should impose maximums and minimums here.
+              "width": 16,
+              "height": 16
+            },
+            "m.alt_text": {
+              "m.text": [{"body": "An image of a 6 sided die with a red question mark over it"}]
+            }
+          }
+        }, {
+          "type": "input",
+          "id": "custom",
+          // Regex the client can use to test input locally. Optional - if not provided the client
+          // should accept *any* input, including an empty string.
+          "validator": "[0-9]+d[0-9]+", // `2d20`, etc
+          "label": {
+            "m.text": [{"body": "Other"}]
+          }
+        }
+      ]
+    }
+  }
+}
+```
+
+In this example, clients which don't support the mixin will see the old-style `!roll 2d6` help text,
+allowing the user to continue interacting if needed. Over time, bots may wish to drop this fallback
+style and instead use a message like `Hello! Your client doesn't support talking to me :(`.
+
+Clients which do support `m.prompts` will instead render the `intro` object as the event. It's not
+required that the `intro.type` matches the top level event `type`, though it is considered good
+practice to do so. The `intro` block is primarily intended to allow senders to tailor their message
+for supported clients, as the intent for this proposal is to discourage commands like `!roll` where
+possible.
+
+Prompts SHOULD be rendered in order of the array, and appear below the `intro` rendering. Buttons
+SHOULD be used for `preset` prompts, using the provided `label`, and text inputs with `label` as a
+prefix or placeholder, and validation per `validator`, SHOULD be used for `input` prompts. For
+example:
+
+![](./images/4139-01-dice-bot-welcome.png)
+
+[Codepen](https://codepen.io/turt2live/pen/gOyVvaY) (note: doesn't do validation)
+
+The user is then able to click on one of the buttons or submit text through the `input` option. That
+reply looks as follows:
+
+```jsonc
+{
+  "type": "m.conversation.reply",
+  "sender": "@alice:example.org",
+  "content": {
+    "m.in_reply_to": { // TODO: Change to match Extensible Event replies
+      "event_id": "$previousMessage",
+      "rel_type": "m.thread" // yes, we use threads!
+    },
+    // Whichever option the user clicked is described here in a new content block.
+    "m.used_prompt": {
+      "id": "surprise"
+    },
+    // We then add all the fallback representations. For `preset` prompts, this is typically just
+    // the `label` verbatim. `input` prompts may require some creative editing, like "Other: 2d20".
+    "m.text": [{"body": "🎲❓"}], // fallback for the image
+    "m.file": {
+      "url": "mxc://example.org/abc123"
+    },
+    "m.image_details": {
+      "width": 16,
+      "height": 16
+    },
+    "m.alt_text": {
+      "m.text": [{"body": "An image of a 6 sided die with a red question mark over it"}]
+    }
+  }
+}
+```
+
+The bot can then process this and continue the conversation as needed, using more `m.prompts` mixins
+to get the information it needs from the user. If the bot considers the conversation/thread to be
+complete, it sends an event with no `m.prompts` mixin to the thread. In our example of a dice bot,
+this could be the result of the roll.
+
+Once a user has picked (and sent) a prompt, the client SHOULD disable the user's ability to send
+another. This could be done by hiding all options, or using the HTML `disabled` attribute.
+
+The example dice bot would then start a new conversation by sending a new welcome message, likely
+with different text to feel less mechanical. For example: "What are we rolling next? [1d6] [...]".
+
+It is left as a bot implementation detail to handle multiple responses, responses from descoped
+users, and invalid input. Typically this would be handled by the bot using a threaded reply to the
+sender saying "sorry, you don't have permission to interact here" or "sorry, I didn't catch that.
+[same prompts as original message]".
+
+## Potential issues
+
+TODO
+
+## Alternatives
+
+[MSC3006](https://github.com/matrix-org/matrix-spec-proposals/pull/3006) is very similar to this
+proposal. Instead of starting per-message threads, it defines interactions via a state event. This
+makes MSC3006 more akin to a "conversation starter" replacement, to use this MSC's terminology.
+
+## Security considerations
+
+TODO
+
+## Unstable prefix
+
+While this proposal is not considered stable, clients should use `org.matrix.msc4139.` in place of
+`m.` in all identifiers.
+
+TODO: Language to support usage in room versions without Extensible Events support, similar to
+[MSC3381: Polls](https://github.com/matrix-org/matrix-spec-proposals/blob/main/proposals/3381-polls.md).
+
+## Dependencies
+
+This MSC has no direct dependencies.
diff --git a/proposals/images/4139-01-dice-bot-welcome.png b/proposals/images/4139-01-dice-bot-welcome.png
new file mode 100644
index 00000000000..54a29679de7
Binary files /dev/null and b/proposals/images/4139-01-dice-bot-welcome.png differ