| name | voice-check |
| description | Use when writing, reviewing, or editing ANY public-facing text before it ships — blog articles, project case studies, LinkedIn/X post copy, README or bio text, meta descriptions — or when asked to "voice check", "de-AI", or review a draft for tone. Also use BEFORE claiming any draft is ready for publishing. |
Voice Check
Two tiers. Tier 1 is mechanical and non-negotiable; Tier 2 is judgment. A review that skips Tier 1 will miss em-dashes and recommend banned constructions as "fixes" (measured baseline failure).
Tier 1 — run the lint, fix every BLOCK
node scripts/voice-lint.mjs <files>
node scripts/voice-lint.mjs --social --platform x <file>
node scripts/voice-lint.mjs --social <file>
Fix BLOCK findings by rewriting, never by exempting:
- em-dash → period, comma, colon, or parentheses (house rule is zero in signed prose; commit
f3077e7 cut 180→6). Never suggest an em-dash as a fix for anything.
- banned-word / marketing-ai / influencer-fluff → delete or replace with a plain, specific word.
- Read every WARN and act or consciously keep (say which).
Tier 2 — judgment checklist
Compare against two published anchors: packages/content/articles/ai-memory-is-cosmetic.mdx, packages/content/projects/goblins-game-parlour.mdx.
- Negative parallelism ("it's not X, it's Y" and kin): at most one deliberate instance per piece. LinkedIn algorithmically penalises the construction.
- Every number and anecdote traces to a source (repo, ADR, eval output, daily note). No invented timelines, no false precision — round honestly ("about a second", not "1.1s" from n=3).
- No invented alliterative frameworks ("The Pattern/Impact/Edge"); cite real named ones (SBI, RRF) in plain language.
- A position is taken. Both-sides framing is an RLHF artefact. Honest uncertainty is ADDED where real ("I don't know", "I might just be screaming into the void") — never cut it.
- First person owns failures: "until I ran it again", not "until someone ran it".
- Rule-of-three padding: triplets where the third item is filler get cut to the real two.
- LinkedIn specifics: → arrows for enumerations (house tic, keep), no bullets-as-emoji, no hashtags, no emoji, link inline at the natural CTA point, no engagement-bait closer, ends on a statement or link.
- X specifics: single tweet (no threads), ≤280 weighted.
- Blog specifics: figures not word-numbers, Australian spelling, short sentences with varied length (deliberate fragments are house style — keep them), section headings are flat claims not gimmicks ("The 10-second nap that ate my pageviews" ✓, "The Uncomfortable Truth" ✗).
Output contract
Findings table: severity | location | quoted text | proposed fix. Then one verdict: SHIP / FIX THEN SHIP (list blocking items) / REWRITE. Quote the lint output line count in the verdict. "Looks good" without lint output is not a verdict.
