Katharsis

Katharsis

A Claude Code output style that classifies each message you send and shapes the reply to fit it.

A status check, an approval, a bug report, and a request for a diagnosis each want a different
reply. Claude Code answers all four with the same shape: a paragraph of narration, the answer
somewhere in the middle, and an offer at the end. Katharsis makes the model classify your message
into one of 11 exchange types before it writes, read a guidance file for that type, and shape the
reply to it: what opens the reply, what stays out, and how long it may run.

Same two messages, same model, same sandbox repo, recorded in Claude Code 2.1.283 and sped up. The
pricing tests expect half-up rounding that finance never confirmed, and the user asks: "should we
change the code or the tests?" The default reply runs 423 words, opens with "Neither, yet", and
ends by offering two more tasks. The Katharsis reply runs 438 words and opens with its
recommendation: "Change the code to half-up, and treat it as provisional until finance answers."
The user then says "go with what you recommend". The default spends 449 words on the work and
opens with "Done, but I need to correct something I told you." The Katharsis reply runs 87 words
and opens with "All 10 tests pass."

To see the same exchange on other models: Claude Opus 5.5 ·
Claude Sonnet 5 · Claude Fable 5.1 ·
Claude Fable 5. Every default reply ends its first turn with an
offer of more work, and one Katharsis reply does, on Sonnet 5. Across both turns, Katharsis is
shorter on Opus 5.5, Fable 5.1, and Fable 5, where it runs 224 words against 540, and longer on
Sonnet 5, 498 words against 349. Every reply is stored verbatim in
demo/captures/, and demo/ has the sandbox and the steps to reproduce
them.

What changes in your…