Every “no, shorter” you tell an AI is a judgment with a reason. CloneTaste collects them, distills them into one portable file, and proves with a score that your agent now gets you.
It's sitting in old chats, then lost when the tab closes. CloneTaste reads your Claude Code sessions and chat exports, finds the moments you corrected, rejected or chose, and keeps only the judgment. Never the whole conversation.
A cheap local filter finds candidates. A fast model classifies them. You approve each one, settle the contradictions, and the rest is held out as a test your agent hasn't seen.
Finds corrections, choices and instructions you repeat 3+ times.
clonetaste scanAccept, edit or reject each candidate. You stay the judge.
clonetaste review“You said shorter here, but longer there.” Pick one, with options.
clonetaste challenge80% becomes TASTE.md. 20% is held out as the exam.
clonetaste distillA blind judge scores agents with and without your taste.
clonetaste interviewTASTE.md is plain Markdown with principles, evidence and the forks where your agent should stop and ask. It works directly as a Claude Code skill.
When an agent asked you to pick and you picked, that's the clearest signal there is. Those become strong judgments without a model call.
Asked for “score the alternatives” three times? It moves to Always do, so you never have to ask again.
Taste isn't always consistent. Challenge asks you to settle it, and the answer is kept as evidence.
For each held-out scenario, two agents predict how you'd react. One has your TASTE.md, one doesn't. A judge sees your real reaction and both answers, without knowing which is which.
The difference is what CloneTaste adds. A clear margin means at least 3 points and more than the spread between runs. If your taste doesn't beat the baseline, the report says so, plainly.
Illustrative numbers. Your report shows your own, with the spread across runs, and warns when there are fewer than 10 scenarios: then it's an indication, not proof.
No accounts, no cloud, no telemetry. Everything lives in a folder on your machine. Only the short excerpts being classified go to the model provider you choose.
TASTE.md, judgments and reports in plain files you own and can read.
Emails, phone numbers, keys and your own list of names are scrubbed before anything leaves.
Runs on your Claude subscription through Claude Code, or your own API key. Nothing else.
Teach it once. Check the score.