菜单
Two-layer AI-tell audit of any text — deterministic scanner (lexicon, constructions, stylometry with line numbers) plus a blind semantic judge. Use when the user wants to know whether text reads AI-generated, what specifically gives it away, and how to fix it.
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
Write a social post in one of the roster's voices (or the user's own `self` voice) — platform-native, craft-floored, scanner-verified. Use when the user wants a post, caption, announcement, reply, or any short-form text written as a distinct human voice.
One brief, N posts, N genuinely different people — voice-isolated writers, cross-post overlap detection, per-post verification receipts. Use for launching/promoting something across multiple posts or platforms without the posts sharing a fingerprint.
Rewrite existing text so it reads human — either scrubbed-and-souled in its own register, or fully re-voiced as one of the roster's people. Use when the user has a draft that "sounds like AI" or wants existing copy converted into a specific voice.
Front door and router for the idiolect voice engine. Use when the user mentions idiolect, asks what voices are available or who should write something, wants writing that "doesn't sound like AI", says text "sounds like ChatGPT/AI slop", or asks for human-sounding posts and no more specific idiolect skill has already matched. Also the first-run onboarding.
Build and maintain the user's OWN voice profile from continuously captured writing plus explicit samples — status, build/refresh, add samples, capture on/off. Use when the user wants posts in their own voice, wants to see what's been learned about their style, or wants to manage capture.
Create a new voice profile — from a writing corpus (measured, not vibed), or invented to fill a gap in the roster. Extends the roster beyond the 60+ built-ins. Use when the user wants a new/custom voice, a voice "like these samples", or a voice the roster lacks.
| name | audit |
| description | Two-layer AI-tell audit of any text — deterministic scanner (lexicon, constructions, stylometry with line numbers) plus a blind semantic judge. Use when the user wants to know whether text reads AI-generated, what specifically gives it away, and how to fix it. |
| argument-hint | paste text or give a file path [--platform x] |
Two independent layers, merged into one report. The scanner catches what regex can see; the blind judge catches what only reading can (no stakes, hollow specificity, symmetric enthusiasm, tutorial cadence). Neither layer knows or cares who wrote the text.
ROOT="${CLAUDE_PLUGIN_ROOT:-${CODEX_PLUGIN_ROOT:-$IDIOLECT_ROOT}}"
# unset everywhere (opencode)? ROOT = two directories up from this SKILL.md
IDIO="python3 $ROOT/scripts/idiolect.py"
Save the text to a scratch file. Run $IDIO scan --file <f> --json [--platform <p>].
Spawn the blind judge — the idiolect-auditor agent — passing ONLY: the text (via the scratch file path), the platform if known, and nothing else. No authorship context, no "the user thinks this is AI," no scanner results (independence is the point; two correlated judges are one judge). On Codex, invoke it explicitly: $idiolect-auditor, audit this file: <path>. If subagent spawning is unavailable, do the semantic pass yourself in a deliberately separate step, AFTER writing down the scanner results, using the auditor's rubric (read $ROOT/agents/idiolect-auditor.md).
Merge into the report:
verdict: 62/100 — reads generated (scanner 58 · judge: generated-leaning)
deterministic tells (line-anchored):
L3 "seamlessly integrates" (lexical T1)
L7 its_not_x_its_y construction
— sentence lengths metronomic (CV 0.31; human ≥ 0.5)
semantic tells (judge):
- zero cost or stake attaches to any claim
- enthusiasm is symmetric across all six sentences
human texture present: 1 exact number; no temporal anchors, no first-person actions
top 5 fixes, in order of impact: ...
Frame it honestly, always: this is a linter verdict about how the text reads, not a forensic finding about how it was made. Skilled humans trip tells; edited AI passes. Never output "this was written by AI" — output "these N things make it read generated, here's the fix for each." If the user is auditing someone ELSE's text to accuse them (a student, an employee), say plainly: no tool can prove authorship from style, false-positive rates hit non-native writers hardest (Liang et al. 2023), and this report must not be used as evidence against a person.
If the user wants the fixes applied, hand off to /idiolect:humanize mode A with the report as the worklist.