360 Feedback Synthesis with Claude

Do this before you upload anything, even anonymized material.

1. Go to claude.ai and sign in.
2. Click your account name or avatar, then look for Settings.
3. Open the Privacy section and find the toggle labeled something like "Help improve our AI models."
4. Switch it off.

What you should see: The toggle moves to an off position, confirming new chats will not be used for future model training. Troubleshooting: Anthropic's own privacy center describes this exact setting under Settings, then Privacy. If the wording has shifted slightly by the time you look, search the settings page for "improve" or "training" rather than hunting through every tab.

This is the step that protects both your client and the people who gave feedback about them.

1. Open the raw survey export or your interview notes in a document.
2. Replace every rater's name with a role label instead: "Peer 1," "Direct Report 2," "Manager."
3. Remove or generalize any job title specific enough to identify one person (a company's only VP of Engineering is not anonymous even without a name attached).
4. Cut any verbatim quote that contains a detail traceable to a single rater: a project name only one person worked on, a specific date, a distinctive phrase they use.
5. Save this cleaned version as a separate file. Never upload the original.

What you should see: A document where every comment is attributed to a role, not a person, and nothing in the text points back to a specific individual. Troubleshooting: If you're not confident a passage is fully de-identified, cut it rather than guess. You can always go back to the original by hand for anything Claude's synthesis misses.

1. From the Claude home screen, click New chat.
2. Click the attachment or paperclip icon and select the anonymized file.
3. Wait for the upload to finish before typing your prompt.

What you should see: The file name appears as an attachment above the message box, ready to reference in your prompt.

Don't ask for finished development priorities on the first pass. Extract themes first so you can check them.

1. Select Opus 5 from the model picker for this kind of pattern analysis.
2. Type a prompt asking Claude to identify the most consistent themes across raters, with a rough count of how many raters mentioned each one.
3. Read the response against your own memory of the interviews.

What you should see: A short list of three to five themes, each with a one-line description and a note on how many raters touched on it. Troubleshooting: If a theme surprises you, don't discard it and don't accept it uncritically. Go back to the anonymized document and check whether the underlying comments actually support it.

1. In the same conversation, ask Claude to turn the themes you've confirmed into three ranked development priorities with a short rationale for each.
2. Review the draft against what you know about the client's role and current goals.

What you should see: A short document: three priorities, each with a sentence of supporting rationale drawn from the themes, ready for you to edit into your own voice.

1. Rewrite anything that sounds generic rather than specific to this client.
2. Cross-check each priority against your independent read of the raw (non-anonymized) data one more time, since Claude only ever saw the cleaned version.
3. Confirm the client's assessment consent form actually covers this kind of AI-assisted processing before you rely on the output in a session.

What you should see: A final draft you'd be comfortable discussing with the client, with every claim traceable back to something an actual rater said.