# Fill it

> Deepgram workshops › Build a voice agent › Mission 03: No dead air › Section 2 of 2
> Web version: https://workshops.deepgram.com/voice-agent/no-dead-air/fill-it
> About 30 min (estimate). Check it: `make check 03` (offline, no key).

## Try the prompt instead

**Do:** End the **Prompt** with this line, then tap **Reconnect and apply**:

> Before you call a tool, say one short line like "One second."

_On the web page, this step has the Settings panel; in your own code the same settings are in the start file._

## Delete the dentist

**Do:** Say "Delete the dentist" and compare with your before.

**Listen for:** a line before the wait.

_On the web page, this step has the agent console: talk or type to the agent there, or run the same mission in your own code._

**Check:** It spoke before the tool answered.

## Hard mode: two covered deletes in a row

**Do:** **Hard mode.** Two deletes in one conversation, each covered within 2 seconds.

**Check:** Hard mode: two covered deletes in a row. (Optional.)

_On the web page, this step shows the code panel with the matching code._

## Take it home

Take it home: [Python](https://github.com/deepgram-devs/voice-agent-workshop/blob/main/python/missions/03-no-dead-air/start.py) or [TypeScript](https://github.com/deepgram-devs/voice-agent-workshop/blob/main/typescript/missions/03-no-dead-air/start.ts).

**What you tuned:** the three-second wait, now filled.

**How it works**

A prompt rule works on any model. The check reads the event log for an agent line between the request and the response, and quotes the seconds from the end of your turn to its first words.

The rule lands some of the time, and that's the lesson. A prompt rule is a request; a text model can drop it when the prompt gets long or the call gets busy, and it writes the filler and the call in one pass, so "one second" can land late. If the check says not yet, delete another item; it scans every call.

**Hard mode:** no **Reconnect** between the two. If one goes silent, tighten the rule, tap **Reconnect and apply**, and start the pair again. The check counts them from the event log.

**With {{name.thinking}}.** The **Settings JSON** tab shows whether it's on here. When it's on: take the prompt rule out, switch on **{{name.thinking}}** in Settings, tap **Reconnect and apply**, and delete another item. A thinking model built for spoken turns knows a call is coming because it decided to make one, so it tends to speak first. Compare the seconds with the prompt rule's.

**Three ways in code:** let the model do it with {{name.thinking}} on, inject a line from your code the moment the request arrives, or ask in the prompt. The injected line covers every call on any model; the [Cover the dead air challenge](https://workshops.deepgram.com/todo-voice-agent/challenges#dead-air) builds it. Whichever you pick, an assistant line between request and response is the proof.

The attendee project has the slow delete and the injected line as its TODOs. For a simpler starter, the [todo-voice-agent guide](https://workshops.deepgram.com/todo-voice-agent) builds the same four tools.

**In production**

- **Watch:** the seconds from the end of the caller's turn to the agent's first words, per tool call, and each tool's own run time. Log the time of every function call request and response next to the agent's first line.
- **Start with:** a covering line for any tool that usually takes longer than about a second, and a timeout on every tool, with a spoken answer when it fires.
- **Read:** the [Voice Agent docs](https://developers.deepgram.com/docs/voice-agent) on function calling and the agent's latency events.
