Calibrate

You are fitting an output filter to THIS user. Their tolerance for warmth, praise, hedging, and scaffolding is not yours to assume. Everything below produces evidence first, rules second.

Phase 1: mine the transcripts

Session logs live under ~/.claude/projects/<project-dirs>/*.jsonl. Search user messages for pushback markers:

grep -l . ~/.claude/projects/*/*.jsonl 2>/dev/null | head -50

Then scan for user turns containing (case-insensitive): "stop ", "don't ", "why are you", "just answer", "no fluff", "get to the point", "stop apologizing", "you don't need to", "skip the". For each hit, extract the assistant message immediately before it. That pair (assistant output + user reaction) is a specimen: the user's own past irritation, on the record.

Cap the mining at 20 specimens. If the logs yield fewer than 5, say so and lean on Phase 2.

Phase 1b: bootstrap from existing rules

Before interviewing, read the user's existing CLAUDE.md / global config for informal style rules ("be direct", "no filler", "don't apologize", "terse", "stop hedging"). Each is a pre-specimen: a fight they already had with the model. Present them: "You already have these preferences; formalize them into catalog entries?" Formalized entries get the full pattern + fix + exemption treatment; the informal originals can then be removed to avoid double-instruction drift.