Curating your profile
Edit statements, trace memories back to their source, delete what's wrong, and keep the profile your agent reads accurate.
Extraction gives you a first draft. Curation is what makes it yours.
This matters more than it sounds like it should: the taste panel is the literal text your coding agent will follow. A sloppy statement here becomes sloppy code later, and unlike a bad prompt you'll never see it happen. You'll just get output that's subtly off and won't know why.
Reading the profile
Open the taste panel from the wand button in the top-right, or ⌘K → View Taste Memories. Memories are grouped by category with their evidence shown as small monospace tags underneath each statement.
Read it end to end after a big import session. It takes a minute and it's the only way to catch drift.
Editing a statement
Click any memory to open Edit memory. Rewrite the statement in the textarea and save with the button or ⌘⏎. The evidence that backs it is shown below as a reference while you write.
Edit when a statement is nearly right: the observation holds but the wording is too narrow, too vague, or scoped to the specific screenshot it came from. Rewriting "the pricing table uses tabular figures" into "numeric data is always set in tabular figures" turns a note about one screen into a rule for every screen.
Since your agent reads these verbatim, write them as instructions. Imperative and specific beats descriptive and hedged.
Tracing a memory to its source
Hover a memory and click the source icon. The panel closes and the camera flies to the node the memory came from.
Use this whenever a statement seems wrong before you delete it. Half the time the memory is fine and the reference is the problem: something you pasted to think about rather than to endorse.
Deleting
Hover a memory and click delete. Krafa confirms, because it can't be undone:
Delete this memory? This permanently removes it from your design taste profile. This can't be undone.
Delete anything you'd argue with. There's no cost to a smaller profile and real cost to a wrong one.
Deleting the reference instead
Removing a node from the canvas prunes every memory that came from it. That's usually the better move when a reference shouldn't have been on the board at all. It removes the belief and stops it from being re-derived or turning up as an exemplar image later.
Deleting the memory alone is right when the reference is good but Krafa read the wrong thing into it.
Pruning follows the node, not the statement. If two references independently produced similar memories, deleting one node leaves the other's memory in place, which is correct, but means "I deleted that screenshot, why is it still in my profile" usually has a second source behind it.
A curation pass
Worth doing after any large import, and every few weeks on an active board:
Read every statement
Anything you'd push back on in a design review gets fixed or removed now.
Check the evidence
Values that don't look like something you'd ship mean the statement is built on the wrong reference. Trace it to source and decide whether the node belongs on the board.
Look for over-fitting
Statements naming a specific product, page, or component won't generalize. Rewrite them into the underlying principle.
Look at the empty categories
Nothing under Motion or Voice & Tone means your agent is inventing both. If you care about them, paste references that carry them: easing tokens, onboarding copy, a component library's transition spec.
Verifying what your agent sees
The taste panel and the MCP profile are the same data, but reading it the way your agent does is a useful check. Ask your connected agent:
Read my design taste profile and summarize it back to me in your own words.
If the summary doesn't sound like you, the profile needs work, not the prompt.