| name | voice-critic |
| description | Critique book prose in the author's voice using patterns learned from his own git edit history (not a hand-written rubric). Use when asked to "voice-check", "punch up", "make this sound like me / like Vonnegut / like Cunk", or to review whether a passage is too pompous / cute / self-aware. Catches what the regex scanner (voice-check.py, ~1.4% of real edits) misses.
|
Voice Critic
The author's voice is semantic, not lexical (the most-repeated swap in 34,742
mined edits occurs twice). Regexes can't capture it; exemplars can. This skill
applies the data-driven rubric + exemplar bank distilled from his real edits.
To critique a passage or a changed file
- Read
GUIDES/VOICE_RUBRIC.md (the 11 edit patterns he applies, with examples).
- Read
scripts/voice/voice-exemplars.jsonl (quality-judged real BEFORE->AFTER
pairs — the in-context reward model).
- For the target prose, for each passage output: the offending span, which
pattern (1-11) it trips, and a suggested rewrite in his voice OR
CUT with the
one-line reason it fails the kill-test.
- Flag first: flat captions, pompous/announcing openers, self-aware asides,
defensive-rigor preempts, cliches, and tidy QED bows. Keep the deadpan
wrong-adjective and concrete-absurd lines.
- Discipline (binding): tie goes to the incumbent; do not churn working lines;
no rule-zealotry; keep the author's exact words unless you can name the reason.
Advisory only — flag and suggest; never auto-apply (it ossifies past cuts).
Or run the assembled critic prompt directly:
python scripts/voice/voice-critic.py <file>
python scripts/voice/voice-critic.py <file> --call
Refreshing the model as the author keeps editing
python scripts/voice/mine-voice-edits.py
python scripts/voice/sample-pairs.py
The exemplar bank is quality-judged (a line worth imitating), not merely mined
(better-than-before). Git pairs are worse->better, and better is not good.
