We build and run websites for the people who hire us, and every project involves a lot of talking. Phone calls about what they want. Messages later about the site once it is live.
For years the hard part was what happened to all that talking afterwards. Someone took notes, and the notes sat in a document. By the time anyone read them again, the client's own words had shrunk to a vague line like "update homepage copy." Two tools fixed that for us, one for each kind of conversation.
Every call turns into a written record
We record our client calls with an app called Memory Notch. It runs on a Mac and starts from the little black strip at the top of the screen. It writes out what people say while they are still saying it, and it picks up both sides of a call, so the client's voice and ours end up in the same place.
When the call ends, three things are saved in a folder on the computer: the audio, a written version of the conversation, and a short note. They are ordinary files, much like a photo or a letter. We can open them any time, using the app or anything else on the machine.
An assistant reads the note and sorts it out
We use an AI assistant for the next part. It already helps us build the website, so we give it permission to read that client's folder.
Then we ask it something simple:
- Which parts of this morning's note are changes to the website?
- Which parts are decisions I should write down?
- Show me the exact sentence each one came from.
What comes back is a list where every item quotes the client. That matters, because it lets us check the assistant's work. If a line looks right or wrong, the recording is sitting in the same folder, and we listen back.
Where the privacy line sits
The recording and the written version are both made on the computer itself, so the conversation stays on that machine. That was the main reason we chose this app.
Sending a note onward to an AI assistant is a separate choice, and we make it one client at a time. Once a note reaches an online service, that company's privacy rules apply to it. So the question worth asking is what leaves the computer afterwards, and it is worth settling before the first call.
What the app leaves to us
The app keeps its job deliberately small, and it helps to know where that job ends.
It records the conversation and writes it out, and it leaves the summarising to us.
It labels the speakers by number, so a written conversation reads as Speaker 1 and Speaker 2. Where two people talk over each other, those labels get rough.
Its built-in search looks at the titles and dates of your notes, so finding a particular phrase from inside a conversation is a job for something else.
We picked it with all of that in mind. Once the conversation is an ordinary file we control, summarising and searching become easy to add. Our assistant writes the summary using everything it already knows about that client, and we can search hundreds of meetings at once. The one thing only the app itself could give us is the promise that the recording stays on the machine.
A tool that leaves you ordinary files hands you something you can build on. A tool that keeps everything inside a clever dashboard hands you a dashboard.
What clients say about the site
The other half is feedback once we have built something and the client is looking at it.
This used to be the messy part. A client would open the preview of their new site, and something would feel off. We would get an email titled "few small things" that began "on the page with the team photos." Then we worked backwards, guessing which page, which button, and what "cramped" meant to the person who typed it.
givefeedback.dev handles that now. It adds a small recorder to the site. The client talks while looking at the page. It captures where they click and scroll alongside what they say, plus a replay of the screen as they saw it. So "this button feels wrong" arrives attached to that button, at the moment they were looking at it, in their own voice.
Pauses come through too, which surprised us. Someone scrolls to a section, stops, scrolls back, then says "it's fine." The pause tells us something the words alone miss.
The assistant collects the feedback itself
The feedback tool connects straight to our assistant. We set that up once per project, and after that it can fetch the recordings, the written versions, and the click history whenever we ask.
That removes a chore. Somebody used to open a dashboard, copy a quote out of it, and paste it into a task list. Now we ask the assistant to read the new feedback and write up the ones carrying real work. Each task it writes points back to the moment it came from.
Why the two of them together
Both tools answer the same question at two different moments. Can a machine read this conversation as it already stands, so a person is free to spend the hour on something better? Memory Notch answers it for the call. givefeedback.dev answers it for the site. Everything after that is one assistant doing one job across two kinds of record.
The real risk here is a quiet one. Forgetting what a client said is rare. What happens far more often is that their words get tidied up. They become a summary of a summary, three drafts away from the sentence they actually used, and the team spends a month building against that. Keeping the original wording close to the thing that writes the tasks is the whole trick.
If you run a small business and the journey from conversation to finished work still runs through somebody with a notepad, those are the two places to look. Start with whichever one costs you more.