"Help improve our AI models"
Off. Your conversations shouldn't be training anything, least of all the ones about your clients.
Unlocked
01 · PRIVACY
02 · PRIVACY
03 · CAPABILITIES
01 · THE SCRUBBER
02 · SESSION NOTES
03 · PATTERNS
04 · CONTENT
Three settings to switch off first
Before the scrubber, before anything else. In Settings, under Privacy and Capabilities. Grey, not blue.
"Location metadata"
Off. Claude doesn't need to know where you are to write session notes.
Memory
Off if you handle client information, so nothing from one conversation turns up in another. On a Pro account, also switch off creating memory from sensitive information.
And one habit: don't press thumbs up or thumbs down on a chat with client material in it. That sends the conversation off for review.
Clean in six steps
- Switch off the three settings above. Once. Then check them again whenever the app updates.
- Get the transcript out of your recorder. Fathom, Otter, Zoom or Teams: export or copy the transcript as plain text.
- Scrub it in a plain chat, not a project. Open a fresh chat with no project, no skills and no connectors. Paste prompt 01, then the transcript. That way the raw version never lands anywhere it could be reused.
- Save the mapping table privately, then delete the chat. Copy the table into a file on your own computer, ideally in a password-protected folder. Then delete that chat.
- Use the clean version for the real work. Session notes, patterns across a programme, content ideas. Prompts 02 to 04 are the ones I use most.
- Put the names back yourself. When a note goes back to the client, swap the labels for real names by hand from your mapping table. It takes a minute, and it's the point.
The prompts
01 is the scrubber. Everything after it runs on the clean version only.
Scrub the transcript
In a plain chat, before the transcript goes anywhere else.
I'm going to paste a call transcript. Before you do anything else with it, scrub it: 1. Replace every person's name with a consistent label (Client A, Colleague B, Partner C) and keep a mapping table of label to real name. Include nicknames and first-name-only mentions. 2. Replace company names, product names, project code names and place names with generic labels. 3. Remove email addresses, phone numbers, amounts of money attached to a named person, and anything else that could identify someone. 4. Keep everything else exactly as spoken. Don't summarise, don't tidy the grammar. Return two things: the scrubbed transcript, and the mapping table on its own at the end so I can save it separately and delete it from this chat.
Notes from the clean transcript
Run on the scrubbed version only.
From this scrubbed transcript, write session notes: 1. The three most important things Client A said, in their own words. 2. What they committed to doing, with dates. 3. What they're avoiding or circling. 4. One question I should open the next session with. Keep every label exactly as it is. Under 300 words.
Patterns across a programme
Drop several scrubbed transcripts for the same client into your project folder first.
Read all the scrubbed transcripts for Client A in this folder, in date order. Tell me: what's changed in how they talk about the problem; the words or beliefs that keep coming back; what they said they'd do versus what they did; and the shift they probably haven't noticed yet. Quote them. Keep every label as it is.
Ideas from what clients actually say
The best content is a sentence a client said at minute 42.
From these scrubbed transcripts, pull the ten most striking things clients said: the ones that name a problem sharply, or a belief most people in their position share. For each, suggest one post idea. Leave out anything that could identify the person even with the labels removed: industries, company sizes, places or events specific enough to point to someone.
Downloads
When it's not quite right
It missed a name
It will, occasionally, especially nicknames and names mentioned once. Read the scrubbed version before you use it. Add the missed name to the prompt ("also replace Jonno") and run it again.
Is this enough for GDPR?
I'm not a lawyer, and this isn't legal advice. It's a strong habit. Check that your contracts and privacy notice say how you use AI, and tell your clients. Team and Enterprise plans come with a data processing agreement, which helps.
Can't I just use a project with memory off?
You can. Scrubbing first means the raw transcript never goes into a project at all. Belt and braces.
Why not let the recorder's own AI write the notes?
Some do a decent job. But then the notes, names included, live in another company's tool. Scrubbed notes in your own folder are yours.