voice-check

voice-check

Answers "would I have written this?" with measurements instead of a hunch. A model judging holistically is style-biased and inconsistent; the script compares one draft with the per-message distributions of the user's own register and reports exactly what is off. It never rewrites.

When to use

• "does this sound like me?", "check / score / verify this draft against my voice", "run the voice check", "is this too formal for me?"
• As the gate inside voice-write, and in the setup prompt's verification bracket.

When NOT to use

• Any rewrite ("make it sound like me") → voice-write.
• "does this sound like AI?" in general → an AI-slop linter; this checks against one person's profile, not a generic tell list.
• Grammar, spelling, "is it professional", "is it too long in general" → not voice.
• No profile.json → voice-corpus + voice-profile first. If the user insists on an opinion anyway, give it labelled "unmeasured" — never present it as a check result.
• Recorded or spoken voice, "does this sound friendly", "is this too long in general" → not this skill.

Workflow

1. Save the draft to a file (temp is fine). Decide the register from the recipient and channel; when unsure ask, or pass the closest one — the script falls back to the same channel, then to all-corpus bands, and says so.
2. Run (--register is required: register-scoped bans are judged against the register you ask for, even when statistics fall back to another one):

   python3 "${CLAUDE_SKILL_DIR}/scripts/voice_check.py" draft.md --register email-external \
     [--brief brief.md] [--exemplars picks.json] [--json]
   

   An incompatible or malformed profile.json exits 2 with a one-line error — rebuild it with voice-profile rather than working around it. --brief enables the unbriefed-identifier check (names, dates, numbers, URLs in the draft but not in the brief); --exemplars (JSON from pick_exemplars.py) enables the copied-span check.
3. Report the verdict and the findings, grouped by severity (precedence FAIL > WARN > INCONCLUSIVE > PASS; exit 1 only on FAIL, exit 2 on configuration errors):
   • FAIL: a user-confirmed ban; length > p90 × 1.25; paragraphs > p90 + 1; an opener or sign-off class the user has never used in a register of ≥ 60 samples. Length, paragraph and class failures need a register with ≥ 10 samples and no fallback — otherwise they are reported as WARN.
   • WARN: a band exit on a short list of features (sentence length, contractions, hedges, politeness markers, question/exclamation share, lowercase starts); a rare (< 5 %) or, below 60 samples, never-seen opener/sign-off class; an AI-default phrase absent from the corpus (never-candidate, advisory until confirmed); caricature (a signature starter used twice, two distinct signature starters under 150 words, the same opening 3+ times, a repeated 4-gram); identifiers (names, dates, numbers, URLs) absent from the brief — an identifier check, not fact verification; copied 6-word spans from the samples; a profile older than 180 days.
   • INCONCLUSIVE: under 40 words the distribution checks are skipped; bans, length, classes and leakage still run, and a ban still FAILs. Quoted lines (>) in the draft are ignored; bans match as case-folded whole phrases, so ban phrases, not single common words ("best" would hit ordinary prose).
4. Do not fix the draft. Point at what would change the verdict and stop; the user (or voice-write) edits.
5. When asked whether the warnings can be trusted, run the self-test and read it out:

   python3 "${CLAUDE_SKILL_DIR}/scripts/voice_check.py" --selftest --loo [--register R]
   

   --loo rebuilds each register without the sample being scored (honest for small corpora; without it the numbers are in-sample and only a floor). Report the overall FAIL/WARN/INCONCLUSIVE shares and the top finding types: a WARN type that fires on 20 % of genuine samples is noise for that register; FAIL should be near 0 % on registers with ≥ 30 samples. Short-message registers (chat) come back mostly INCONCLUSIVE — say that coverage is limited there rather than implying a score.

Output spec

VOICE CHECK: <verdict> plus one line per finding with the measured value and the user's band or class shares (63 words — far above your p90 of 29, closes with "best regards" — 0 % of your external mail; you use: best 38 %, thanks 33 %, name-only 28 %). --json gives {verdict, register, fallback, flags[{check, severity, message, ...}]}.

Gotchas

• Thresholds are repo defaults, not universals; only the self-test tells you how they behave on this corpus. The bands are per-message p10–p90, so a genuine message lands outside one band roughly one time in five and outside some band more often — that is why band WARNs are advisory and only bans, length, paragraphs and never-seen classes can FAIL.
• Class checks need a register with ≥ 20 samples (WARN) and ≥ 60 to FAIL — with 0 of 60 the true rate is below ~5 % (rule of three); with fewer, absence proves little. The default home is ~/.personal-tone; pass --home for anything else so a shared working directory can never supply someone else's profile.
• Lexicon features (contractions, hedges, politeness, AI vocabulary) are English; the script skips them for registers the profile marks unreliable.
• Not an AI detector. Detectors false-flag non-native writers at ~61 % (Liang et al., Patterns 2023); this tool compares against the user's own baseline only.
• Needs voice-profile/scripts/stylometry.py next to this skill (same repo install) or PERSONAL_TONE_STYLOMETRY pointing at it.

Files

• scripts/voice_check.py — the checker (exit 1 on FAIL) and --selftest.
• scripts/test_voice_check.py — golden set: a genuine sample passes, a generic assistant rewrite fails, a caricature and a wrong-register opener are caught, copied spans and unbriefed names are flagged.
• references/checks.md — every check with its threshold, severity, rationale and evidence, plus the features deliberately excluded as noise on short text.