How to Build a Live AI Sales Coach with Claude Code

What is a live AI sales coach?

A live AI sales coach is a small tool that sits on your screen during a call. It listens to what the prospect says and what you say, compares the conversation with your sales script, and tells you in real time what to do next. It can flag when you drift off the script and fill in sections such as the prospect's constraints as the call goes on.

For an agency owner who is not a natural closer, or who has several closers to train, it gives live feedback without anyone sitting in on every call. This build runs locally on a laptop. It is a proof of concept, not a deployed product.

How do you start the build in Claude Code?

Open Claude Code, select an empty project folder and switch to plan mode. Claude Code has several modes: ask permissions, accept edits, plan mode, auto mode and bypass permissions. When you begin a new tool, use plan mode, so it plans before it does anything.

Our first prompt was short and did not specify any technical set up. It said we wanted a real time sales coach that uses audio capture to listen to calls, gives feedback against a script, shows prompts in a small overlay window, flags missed points, detects different speakers, runs locally and stays simple. We dictated it with the voice input in Claude Code, then told it to ask as many questions as it needed.

What decisions does Claude Code ask you to make?

Plan mode will ask a long list of questions. The answers we gave shaped the tool:

• Audio capture: system audio plus the microphone.
• Interface: a small floating overlay on the laptop, not a browser tool.
• Speech to text: cloud streaming, because it is quicker than running Whisper locally.
• Coaching model: Haiku, which can be changed later.
• Script format: one flat playbook, not separate stages for discovery, demo and close.
• Live prompts: the next step from the script, objection handling cues and missed opportunities.
• Suggestions: one main suggestion with a small history, with a big start button.
• Saving: transcripts and the coaching log saved to a local file, with a context box to paste details about the prospect before each call.

Why ask for more questions before accepting the plan?

In plan mode, Claude Code can rush to a plan so it can start building. Before approving, we asked it for ten more questions so it had full context. This is not always needed on simple projects, but it is a cheap way to make sure the tool matches what you want.

Read the plan once it appears. The main thing to check is that it has understood what you are building. When it looked right, we accepted it in auto mode so it would not stop to ask permission every few minutes.

How do you set up the keys and the script?

The first version needs a few things from you before it can run. You can ask Claude Code to give you step by step instructions for each of them.

1. Create an account with a speech to text provider and generate an API key for transcription. We used Deepgram.
2. Create an API key in the Anthropic console for the coaching model.
3. Put both keys in a local environment file, created from the example file Claude Code made.
4. Paste your sales script into the script file in the project.
5. Ask Claude Code to check that both keys work, then start the app locally.

What goes wrong, and how do you fix it?

The first version ran but was hard to read and would not start listening. The fix was to keep giving specific feedback and let Claude Code correct itself. Our rounds went like this:

• The overlay was too transparent and too vertical. We asked for a toggle to hide or show the transparency and a horizontal bar at the top of the screen.
• Pressing start failed with a mic capture error. Claude Code checked the logs and suggested closing other apps using the mic and checking the Windows privacy settings. We asked it to find a way around the limit, because on a real call Zoom or Google Meet will already be using the microphone.
• The coach picked up the system audio but not the microphone. We asked for a live transcript and an audio level meter, so we could see what it heard, and then for a microphone selector. After that, the mic worked.
• Suggestions that appear while you are speaking can distract you, so the coach needs to hear both sides of the call to stay relevant.
• We asked for the suggestion to turn green when we said what the script asked for, then move on to the next one, plus a preview of what comes next. We also asked it to show the actual opener instead of just the words run the opener.

How do you test a sales coach before a real call?

First, play back a recorded sales call through the system audio and watch what the coach suggests. Then run a role play on your own, where you speak the script yourself. Finally, ask someone to play a prospect on a video call while you run the coach in the background.

The live tests showed that the coach could pick up what was said, give useful prompts, and move on when we covered a point. They also showed problems. It sometimes got stuck on a section even after we had covered it, and the skip button did not work at first. When we jumped to a later section, the high priority prompt did not change. At one point the section list showed the word object instead of names, which Claude Code fixed quickly.

To fix these we asked for three changes: a skip button with a drop down to jump to any step, a high priority prompt that updates when you change section, and a context panel on the right that builds up key points about the prospect, such as their constraints. We also shrank the transcript once we knew it worked and moved the mic and system status to the top bar.

What can you use a call coach for besides sales?

The coach follows whichever script you load, so the same tool can support client calls, upsell calls and team training calls. If you train a team of closers, it gives each of them live feedback without you joining every call.

The final version was still not perfect. It threw a lot of information at the screen, which can trip you up on a real call, so it needs fine tuning over time. As a proof of concept it showed the main idea works: a coach that listens, follows your script and tracks what the prospect says.