voice-check

Voice Check

Table of Contents

• Workflow
• Don't-list regex bank
• Worked example
• Guardrails

Related skills: Called by the Editor after structural-review. Runs alongside hedge-detector, slop-detector, citation-form-check. Each skill owns a specific band of voice issues; this one handles explicit don't-list phrase matches.

Workflow

Voice check draft D:
- [ ] Step 1: Load voice-profile.md (global) + voices/{section}.md (if applicable)
- [ ] Step 2: Extract the don't-list as a regex bank
- [ ] Step 3: Scan draft; match each regex with word boundaries
- [ ] Step 4: For each match: record {line, quoted phrase, violation, voice-profile citation, 2 rewrite options}
- [ ] Step 5: Emit phrase-flags table

Don't-list regex bank

From voice-profile.md §10 (Voice don'ts). Word-boundary enforced.

Violation	Regex	Rewrite hint
delve	`\bdelve(s	d
unpack	`\bunpack(s	ed
dive into	\bdive\s+into\b	replace with specific verb
let's explore	`\blet'?s\s+(explore	dive)`
at the end of the day	\bat\s+the\s+end\s+of\s+the\s+day\b	delete clause
game-changer	\bgame[-\s]?changer\b	name the specific change
paradigm shift	\bparadigm\s+shift\b	name the shift explicitly
under the hood	\bunder\s+the\s+hood\b	replace with "mechanically" or delete
in today's fast-paced world	in\s+today'?s\s+(fast[-\s]?paced\s+)?world	delete; open with a dated concrete fact
AI is transforming	(?i)AI\s+is\s+transforming	delete generic framing; open with confession
emoji	[\u{1F300}-\u{1FAFF}\u{2600}-\u{27BF}]	delete
exclamation	! (in body, not quotes)	convert to period
I think (as hedge)	\bI\s+think\b (when followed by a claim, not a question)	commit OR specific hedge
clearly / obviously / simply (as persuasion)	`\b(clearly	obviously
custom CTA	`subscribe\b.*(resonated	more

Section overlays can add or override entries. If voices/kalshi-log.md bans "generalizing to other sports", add that rule.

Worked example

Draft excerpt:

In today's rapidly evolving AI landscape, let's unpack why RAG beats fine-tuning. At the end of the day, RAG is a game-changer. Let's dive into the details!

Flags (all tier-1):

loc	quote	violation	citation	rewrites
P1 S1	"In today's rapidly evolving AI landscape"	generic opener	voice-profile §10 don't #2	(a) delete; (b) replace with a dated concrete fact
P1 S1	"let's unpack"	don't-list	§10 don't #1	(a) "Here is the mechanism" (b) delete
P1 S2	"At the end of the day"	don't-list	§10 don't #1	(a) delete clause
P1 S2	"game-changer"	don't-list	§10 don't #1	(a) "it cuts token budget by 40% on our traces" (b) "it avoids a retrain"
P1 S3	"Let's dive into"	don't-list	§10 don't #1	(a) "Here is what actually happens" (b) delete
P1 S3	"!" (exclamation)	don't-list	§10 don't #8	(a) period

Six tier-1 flags in three sentences → triggers Editor must-not #13 (≥3 tier-1 voice violations = no-go).

Guardrails

1. Every flag must cite a voice-profile line. No flag without citation.
2. Regex must have word boundaries; avoid false positives (e.g., "undelete" shouldn't match delve).
3. Section overlays extend but never override global don'ts — an overlay cannot allow a global don't.
4. Exclamation-point rule excludes quoted dialogue and code fences.
5. "I think" is flagged only when followed by a claim. "I think so, but I'm not sure" is a sentence of quiet agreement, not a hedge on a claim.
6. Max 2 rewrite options per flag.

Quick reference

• Input: draft + voice-profile + optional overlay.
• Output: phrase-flags table in the voice pass.
• Complements hedge-detector (hedging) and slop-detector (structural voice).