# How to Build an AI Meeting Coach with Claude Code

> An AI meeting coach takes a call transcript, detects whether it was a sales, customer success, onboarding or team call, scores it against your own framework, and turns the commitments made on the call into action items. You can build a working proof of concept with Claude Code by describing the problem in plan mode, adding an AI API key, connecting a task database such as Notion, and uploading your frameworks.

Source: https://fluxomate.com/blog/how-to-build-an-ai-meeting-coach-with-claude-code.html  
Published by Fluxomate (https://fluxomate.com/), 2026-06-18. Section: Claude Code builds.
Video: https://www.youtube.com/watch?v=LjfV8DWqbQ8

## Key takeaways

- Most agency calls never feed back into coaching. A meeting coach gives every call a score, feedback and a list of commitments.
- The tool detects the call type, then grades it against the framework you uploaded for that type of call.
- Action items are extracted from the commitments made on the call and sent to your task tool, with context about the call attached.
- Connecting an existing tool such as Notion needs an access token and a database link, both stored in a local environment file.
- Upload your real frameworks and SOPs. Without them the coaching is generic.
- A tool people have to paste transcripts into will not get used. Plan to connect it to your meeting recorder so it runs by itself.

## What does an AI meeting coach do?

An agency can run hundreds of meetings a month, and only a small share of them are ever used to coach the team. A meeting coach changes that by reviewing every call and giving the person who took it specific feedback.

In this build, you paste a call transcript into a simple app. The tool then works through four jobs.

The example agency in the build is a Facebook ads agency with around twenty team members, which wants feedback on every call so that nothing slips through the cracks.

Name the problem before you build. Here it is that not enough meetings get coaching, so call quality drops. Pick the problem first, then the solution.

- Detect the meeting type: sales call, customer success call, onboarding call or team call.
- Grade the call against the framework and SOPs you uploaded for that type of call, with a score and written feedback.
- Pull out every commitment made on the call and turn it into an action item.
- Push the action items you choose into a project management tool, so people are held accountable.

## How to write the first prompt

Open Claude Code, choose a new folder on your computer and keep everything for the project in it. Switch to plan mode, which makes Claude Code plan before it creates anything. Planning matters because the more of your thinking you put into the chat at the start, the easier the build is later.

Our first prompt described the tool in plain language and added the context of the agency. It said that users paste in a transcript, that the tool should detect the meeting type, give feedback based on the uploaded frameworks and SOPs, create action items from the commitments made to clients or prospects, and send them to a project management tool. It asked for a simple app hosted locally with a unique, engaging interface, and asked Claude Code to brainstorm and ask questions before building.


## Answering the planning questions

Claude Code replied with a set of multiple choice questions. We took the recommended option for how the app should be built and run locally. Our other answers set the shape of the product.

Read the plan for intent rather than for technical detail. You are checking that it understood what you want, not reviewing the code.

- Action items should land in Notion, to show how easily an internal tool connects to software your team already uses.
- SOPs and frameworks would be dropped in as folders, covering sales, customer success, team and onboarding meetings.
- A Notion database already existed, so the tool only needed to push to it.
- The AI analysis would use an Anthropic API key.
- For the standout features we chose all of them: an animated scorecard, an AI coach persona with feedback, highlighted key moments and a team leaderboard.

## Connecting the AI and Notion

You do not have to use Notion. ClickUp, Monday, Airtable or any tool with an API works the same way. You also do not have to use Anthropic for the AI. Other providers work too.

You can simply ask Claude Code how to get each key and it will give you a guide. The steps below are the ones we followed.

1. Create an API key in your AI provider's console and copy it.
2. In Notion, create a new connection for your workspace and give it permission to read and update content. Copy the access token.
3. Open your task database as a full page, add the connection to it, and copy the database link.
4. Open the environment file that Claude Code created, which is a local file for passwords, and paste in the AI key, the Notion token and the database link.
5. Tell Claude Code you pasted the whole database link, and ask it to extract the database ID. It noticed the new keys on its own and fixed the ID.

## Keeping your keys safe

The environment file is the project's password store, and its contents should stay on your computer. If you publish the code to somewhere like GitHub, this file is not uploaded with it. In a published setup, the keys go in separate settings instead.

You can paste keys straight into the chat, but then they pass through the AI model and could be leaked. Safer is to open the file yourself, paste the key, and save. Never share these keys publicly.


## Testing the coach and fixing errors

The first version looked generic, so we gave Claude Code some interface references and asked for a hand drawn graphite look. That is optional, but it shows how easily you can restyle an interface.

Then we tested the function. The first attempt returned an error saying the API key was not set, even though we had added it. Errors at this stage are normal. We pasted the error into Claude Code, asked it to check the environment file and restart the server, and it worked on a later attempt.

Once working, the app showed a scorecard with a grade and a confidence level, the coach's feedback, strengths, what to fix next, what a ten out of ten answer would have said, a highlight reel of key moments and a list of commitments. We selected two action items and sent them to Notion, where they appeared as tasks with an assignee.


## Adding your own frameworks and improving it

A coach is only as useful as the playbook behind it. We generated example client success, onboarding, sales and team call frameworks, dropped them into the chat and told Claude Code these were the frameworks to follow. After that, the same transcript produced a differently structured review, and the tool correctly identified an onboarding call.

We also asked for two fixes: add more context about the call to each Notion task, so someone reading it later knows what it came from, and make leaderboard entries clickable so you can reopen a past review. Both worked.

- Test with a long transcript. A call of around fifty minutes still worked.
- Check the progress steps. In our build the loading stages moved too fast to be real, so a live launch would need them to reflect actual progress.
- Check the action items. The tool sometimes generated more commitments than we wanted, which is why the user selects which to send.

## Making it something your team will use

Pasting transcripts by hand will not last. The more effort a tool needs, the less it gets used. The better setup connects the tool to your meeting recorder, so a new transcript triggers the analysis automatically and the person who took the call receives their score and feedback in Slack.

Beyond that you can add reminders, analytics, a chat for follow up questions, a memory of each person's earlier calls so it can point out repeated mistakes, and manager views. The video above walks through the full build.


## Questions people ask

### Can I build an AI meeting coach without coding?

Yes. This build used plain language prompts in Claude Code and no hand written code. You describe the tool, answer its questions, add your keys, and give feedback after testing.

### How does the coach know what type of call it is?

The tool reads the pasted transcript and detects whether it is a sales call, a customer success call, an onboarding call or a team call. It then grades the call against the framework you uploaded for that type.

### Do I have to use Notion?

No. Notion was used as the example task tool. Any tool with an API can receive the action items, such as ClickUp, Monday or Airtable, and the setup follows the same pattern of an access token and a database or project link.

### Why connect it to a meeting recorder?

If your team has to paste transcripts in manually, they are unlikely to use the tool. Connecting it to your recorder means each new transcript triggers the analysis automatically, and feedback can be sent to the person who took the call.

### How is this different from using a meeting recorder on its own?

A custom tool can be shaped to your frameworks and workflow. You can add reminders, analytics, a chat for follow up coaching, a memory of earlier calls and manager views, none of which come as standard.

Tools used: Claude Code, Anthropic API, Notion

Free templates for this guide: https://fluxomate.com/resources.html