> ## Documentation Index
> Fetch the complete documentation index at: https://docs.convonexai.tech/llms.txt
> Use this file to discover all available pages before exploring further.

# Talking to the agent

> Chatting, answering the agent's questions, stopping a turn, and tuning how it thinks.

The conversation panel is where everything starts. Tell the agent what to
build, change, or ask about — it keeps context across turns, and it asks for
clarification when it genuinely needs a decision from you rather than guessing.

A few useful habits:

* **Plain English works.** "Make the header sticky and move the login button
  right" is a complete instruction.
* **Technical vocabulary works too.** Say "use Postgres" or "make it SSR" and
  the agent applies your choice rather than debating it.
* **Attach images in chat** — a sketch, a screenshot of a bug, a design you
  like. They're delivered to the agent with your message.
* **Small questions get direct answers.** Asking "why is the total wrong?"
  doesn't trigger a rebuild — the agent answers, and only builds when you ask
  for a change.

## When the agent asks *you* something

Some decisions belong to you — and for those, the agent posts a **question
card** in the conversation instead of guessing. If the question is blocking the
build, the build holds until you answer; the workspace's attention bar surfaces
it on every tab so you can't miss it.

Questions are durable: they're written down before they're shown, and your
answer is recorded before the UI claims success — so a refresh or a dropped
connection never loses either side. When you answer a blocking question, the
card tells you honestly whether the build resumed or is waiting idle.

## Stopping

Press **Stop** in the top bar (or `Esc`) to halt a running turn. The button
shows "Stopping…" until every part of the agent's work has actually quieted —
not just the visible message. Your project is left in a consistent state and
you can pick up with a new instruction.

## Tuning the agent

* **Model** — change the project's model any time from the top bar; the choice
  persists.
* **Thinking dial** — **Fast / Auto / Thorough** controls how hard the agent
  reasons per turn. Auto (the default) is quality-first; Fast is cheapest and
  quickest for simple or throwaway iterations; Thorough applies maximum
  reasoning to every turn. Changes apply from the next turn. For a single hard
  problem, include the word `ultrathink` in your message to escalate just that
  turn.

## Developer-tier extras

On the Developer tier, two more surfaces open up:

* **Memory Lens** — see what the agent remembers about your preferences and
  project facts across turns; peek, pin, edit, or delete what it's carrying.
* **Skill Shelf** — browse the skills the agent can bring to bear.

Switch tiers any time from the top bar — see [The workspace](/building/workspace).