Poe Follow-Up Actions

Overview

Existing Flow

In single-user chat, follow-up actions appeared beneath the latest bot response. Tapping one would trigger another response or compare the answer against a different bot. The resulting interaction appeared as plain text followed by the new bot response.

Key limitations: actions were tied only to the latest response, and the resulting UI was visually disconnected from the rest of the chat.

Problem

In group chats, several issues emerged immediately:

• Should actions appear for everyone or only the triggering user?
• The UI consumed too much vertical space
• Interactions didn't scale visually in active conversations
• The resulting messages felt disconnected from the chat itself

We still needed to support conversation continuity, bot discovery, and cross-model comparison. And there was a second issue: even in single-user chats, follow-up actions only worked on the latest bot message — which felt unnecessarily limiting.

Constraints

Before diving in, here's where we ultimately landed:

To get there, I distilled the requirements into core goals. The system needed to:

• Show who triggered an action
• Show what the action references
• Create a clear interaction entry point
• Indicate that referenced content was tappable

While also:

• Minimizing height
• Minimizing redundancy
• Relying on familiar interaction patterns

The challenge wasn't just visual design — it was designing something users could intuitively understand without explanation.

Entry Point

My first explorations focused on reducing the size of the existing follow-up action UI while preserving the same interaction model.

Even reduced, they still consumed too much vertical space for group chat contexts. So I shifted directions.

Exploration: Nested Overflow Menus

I explored moving follow-up actions into the existing overflow menu — tap overflow, select "Compare," choose another bot from a nested menu.

This tested well internally from a UX perspective. But after discussing implementation with engineering, I learned our existing overflow component didn't support nested interactions — and supporting this behavior would significantly increase scope.

Final Direction

Instead of nesting actions inside the menu, we simplified: the overflow menu becomes the entry point, and selecting it opens a modal containing available actions.

This preserved discoverability while dramatically reducing UI noise in the main conversation surface.

Chat Alignment

Previously, follow-up actions rendered as plain text followed by a bot response.

Early Exploration: Making Complexity Explicit

My initial instinct was to make the system fully legible — rich metadata, explicit labels, heavily annotated UI.

I compared

Someone else compared

Someone else compared

Someone else compared

But these directions violated the original goals of simple, lightweight, and intuitive. The UI explained too much.

Exploration: A Distinct Message Type

Next, I explored making follow-up actions visually unique from other chat messages.

Compare

Image upscale

Voice request

Animation request

Feedback surfaced an important issue: this introduced yet another message type into the design system, increasing both design and engineering complexity. At the same time, we were already hearing that Poe felt overly complex. This direction amplified that problem rather than solving it.

Unifying Decision

I realized follow-up actions and replies could share the same visual and behavioral model. I borrowed from familiar messaging patterns — specifically lightweight reply previews similar to WhatsApp.

Compare (Text)

Image upscale

Voice request

Animation request

1 — Reduced Learning Cost

Users already understood that replies reference another message and that tapping them jumps back to context. We didn't need to teach a new interaction model.

2 — Reduced System Complexity

Replies were already on the roadmap. By framing follow-up actions and replies as the same system, we accelerated reply functionality, reused infrastructure, and justified investing more deeply in the interaction model. This reframing significantly improved cross-functional alignment.

AI Challenges

Challenge 1 — Compare on a compare on a compare

One particularly tricky issue emerged around comparison chains. "Compare" lets users generate responses from different bots to the same prompt — GPT-5, then Claude, then Gemini. Compare works by tracing back to the original user prompt, stripping bot mentions, and excluding previous bot responses from context.

Option 2 matched reply mental models — previews show what you clicked — but it misrepresented what the model was actually seeing. Behind the scenes, compare always traced back to the original user message. I pushed for Option 1 because it aligned with a core product principle:

Challenge 2 — The Context Window Problem

Models also silently lose access to older messages as context windows fill — even when users can still visually see them. This created edge cases affecting follow-up actions, compare behavior, and referenced message navigation.

History visible but not accessible

Context cleared

Failure States

We considered rehydrating older messages and rebuilding context dynamically. Technical limitations made this difficult. Instead, I pushed for simple, transparent, honest failure states — rather than hiding AI limitations, we exposed them clearly and predictably.

No follow up actions

Reference can't be accessed

To validate whether deeper infrastructure investment was necessary, I partnered with engineering to add logging around these failure states post-launch.

Results

We landed on a system where replies and follow-up actions share one interaction model, the UI scales naturally in group conversations, interactions remain lightweight and familiar, and AI-specific edge cases are handled transparently.

A major implementation win: this work effectively enabled two roadmap projects simultaneously — the follow-up actions redesign and reply infrastructure.