Messages & Responses

This page covers everything about the messages themselves — how PebbleChat thinks before it answers, how it streams its response in real time, how to export results, and how to iterate when the first answer isn’t quite right.

How PebbleChat produces a response

When you send a message, PebbleChat doesn’t just feed your prompt to a model and stream back the output. It goes through several phases you can watch happen in real time:

1. Planning — the model thinks about how to approach your request. For reasoning-capable models (Claude Opus, o1, etc.) you can expand this into a full reasoning trace.
2. Research / Tool use — if the task benefits from fresh data, PebbleChat runs web searches, calls MCP tools, queries document stores, or invokes @mentioned flows. This happens inside the Activity Stream (see below).
3. Finalising — PebbleChat synthesises everything it gathered into a single cohesive response.
4. Response — the final answer streams in as formatted markdown.

You see all four phases play out on screen. The Advanced Features page covers the Activity Stream and thinking panels in depth.

Watching the model think

For complex tasks, PebbleChat shows a Research & Tools activity banner above the final response. Expand it and you’ll see:

• Each research step the model took (web search queries, tool calls, document lookups)
• The sources it consulted — clickable links to what it found
• Asset discovery hits — any flows, agents, or document stores PebbleChat pulled in automatically
• Per-step timing — how long each phase took

For reasoning-capable models, PebbleChat also shows a Thinking panel with the model’s chain-of-thought reasoning before the final answer. This is surfaced automatically for models that support extended thinking; you don’t need to configure it.

Why this matters:

• You can catch reasoning errors early — if the model is about to go down a wrong path, you’ll see it in the plan before the final answer is committed
• You see what sources were consulted — critical for trusting the response on research-heavy tasks
• You learn how the model approaches things — useful for prompt engineering and for coaching it next time

Activity Stream collapses automatically once the response is complete, with a summary at the top you can re-expand any time.

Real-time markdown streaming

Responses render progressively as markdown while they stream. You don’t have to wait for the whole response to arrive before you can start reading — headings, bullet lists, tables, code blocks and links all appear as they’re generated.

Features you’ll see streaming live:

• Formatted headings and nested lists
• Tables with full borders and alignment
• Code blocks with syntax highlighting (a copy button appears on each block when complete)
• Inline links to sources
• Bold, italics, blockquotes
• LaTeX for mathematical expressions (when the model produces it)

Auto-scroll keeps the latest content in view. If you scroll up to read something, auto-scroll pauses so you aren’t yanked back down — scroll to the bottom yourself when you’re ready.

Stopping a response mid-stream

If a response is going in the wrong direction, click Stop (replaces the Send button while a response is streaming). The response halts at whatever it had produced so far, and you can:

• Edit your prompt and resend
• Start a new sibling message without the partial response affecting context
• Leave it as-is if the partial output is useful

Message actions

Hover any AI response (or your own message) and a set of action icons appears:

Exporting a response

PebbleChat can export any response into several document formats. Click the export icon and choose:

Note on complex content: Charts, rich Crayon components, and some interactive visualisations can’t be faithfully rendered in every export format. Tables and text always export cleanly. Check the exported document before sharing.

Exports include the message body only — by default they don’t include the conversation context, your original prompt, or the activity stream. This keeps exports clean for sharing.

How to formulate good messages

A few patterns that produce noticeably better responses:

Be specific about the output shape

Vague: “Help me with the pricing page” Better: “Rewrite the pricing page copy in a more confident tone. Keep the existing three-tier structure, but update the value propositions to emphasise time saved rather than features.”

Include your role and the audience

Vague: “Explain OAuth” Better: “I’m briefing a non-technical product manager on why we need to upgrade from OAuth 2.0 to OIDC. Give me a 3-paragraph explanation focused on user experience and risk, not cryptographic details.”

Show what you’ve already tried

Vague: “The query is slow” Better: “I have a Postgres query joining three tables, 10M rows. Current execution time is 8 seconds. I’ve added indexes on the join columns and run ANALYZE. What should I investigate next?”

Ask for structured output when it helps

“Respond as a table with columns: Problem, Impact, Recommendation” “Give me 5 bullet points, each with a heading in bold and a one-sentence explanation” “Return a JSON object with keys: summary, risks, next_steps”

Referring to past messages

You can scroll back to any earlier message in the conversation to re-read it. Clicking an earlier message doesn’t “rewind” the conversation — it just shows you that message. The model still has the full conversation context when you send your next prompt.

To actually branch a conversation from an earlier point — keeping everything up to that point but abandoning the later exchanges — edit the earlier message and resend. PebbleChat treats the edit as a new branch and re-runs from that point.

Iterating on a response

Common patterns for refining:

• “Make it shorter” — usually works in one step
• “Rewrite that in a more formal tone” — tone adjustments are a strength
• “Explain your reasoning for point 3” — drilling deeper into a specific part
• “Give me three alternatives” — asking for options instead of a single answer
• “That’s wrong because X — try again” — corrections with a reason work better than just “wrong”
• “Format that as a table” — post-hoc formatting adjustments

What PebbleChat does not have

A common question: “Where’s the thumbs up / thumbs down?”

PebbleChat does not have per-message thumbs-up / thumbs-down feedback today. What it has instead is the Submit Feedback modal in the help menu, which is draggable, supports screenshots, and is wired to a Jira ticket the engineering team triages.

If a response is wrong, worth correcting, or worth celebrating — use Submit Feedback. It gives the team much more actionable information than a thumbs rating.

Related

• Advanced Features — Activity Stream, thinking panels, and the context window indicator in depth
• @Mentions & Tools — how the research phase uses your org’s flows and document stores
• Model Selection — how picking a different model changes the thinking/response behaviour