authenticity-check

A standalone, pure-prompt skill that scores how authentically a piece of text reads as the work of a real human author, and flags the specific spans that read as AI-generated, AI-templated, or generically derivative. It is just instructions: SKILL.md plus a few reference files. No scripts, no dependencies, no network access. Its tools are read-only by design.

It is the evaluative counterpart to the humanizer skill. This one diagnoses. It does not rewrite.

Where this comes from

This skill is the diagnostic half of a pair. Its detection criteria descend from the voice-preservation logic that powers Scriveno (formerly Scriven), an AI-native longform writing, publishing, and translation pipeline whose core promise is that drafted prose should sound like the writer, not like AI. humanizer lifted the de-slop, restraint, and voice-matching layer of that pipeline into a standalone rewrite skill. authenticity-check is the read-only counterpart: it applies the same catalog and the same restraint to diagnose rather than transform, and it vendors humanizer's criteria so the two agree on what a tell is. If you want the rewrite, use humanizer. If you want the whole writing pipeline, see Scriveno (npm package: scriveno-cli).

What it does

Given a piece of text, it returns:

• an authenticity band (Reads human / Mixed signals / Reads AI-generated),
• a 0-100 authenticity score (higher means it reads more like a person's own work),
• span-level flags, each with the reason it reads as AI-generated, AI-templated, or derivative,
• a "Reads as human" section naming what it deliberately did not flag,
• and a caveat that the score is a heuristic read, not proof.

It runs in generic mode by default, and in a voice-deviation mode when you ask whether a draft still sounds like you (or a named author) and a voice sample or profile is available.

The pairing with humanizer (a pair, never a merge)

authenticity-check and humanizer are designed to be used together, as separate skills, with a human deciding between them:

authenticity-check  ->  you read the flags and decide  ->  humanizer
   (diagnose)               (judgment, not automation)       (rewrite)
        ^                                                        |
        |________________ re-verify, fresh read _________________|

They are deliberately kept apart. A single tool that scores text and then rewrites it to raise its own score is a detector-gaming loop: it optimizes prose against a metric instead of improving it for a reader. The humanizer skill explicitly refuses that loop. Keeping diagnosis and transformation in two skills, with a person in between, is what prevents the loop from forming. So this skill never rewrites, never carries a "target score" into a rewrite, and never instructs humanizer how to move a number. The re-verify step is a fresh diagnosis of the revised text, not a continuation of the first.

voiceprint (future bundle, designed for, not built here)

A future product, voiceprint, will bundle humanizer (transform) and authenticity-check (diagnose) into one combined offering. This repo is built to compose cleanly into that bundle: no assumption that it is the only skill installed, clean separation of the skill from its criteria files, and no global names that collide with humanizer (the skill name, the Cursor rule filename, and the frontmatter name are all authenticity-check). The band vocabulary is chosen to sit alongside humanizer's density language so the two reports compose without a translation layer. voiceprint is not built here; this repo simply does not block it.

Vendored criteria and the sync obligation

The detection criteria live canonically in the humanizer repo (references/tell-patterns.md, references/do-not-flag.md, and references/voice-matching.md). Because authenticity-check is a standalone repo and the humanizer repo must stay untouched, this repo vendors synced copies of those three files into its own references/ directory. Each vendored file carries a header stamp recording that it is a synced copy, that its canonical upstream is the humanizer repo, and the obligations below.

• These three files are synced copies, not the source of truth: references/tell-patterns.md, references/do-not-flag.md, references/voice-matching.md. Last synced 2026-05-29 from humanizer commit 9632cf1. The stamp records the last criteria sync, not every humanizer commit; humanizer changes that do not touch the three vendored files (for example, adapter additions) do not require a re-vendor.
• Re-sync when humanizer's criteria change. Do not edit the criteria in this repo independently. A fix belongs upstream in humanizer and is then re-vendored here. Editing the copies in place makes the two repos drift and the diagnose/rewrite pair stop agreeing on what a tell is.
• Reconciliation is planned. When the voiceprint product is built, these vendored copies and humanizer's originals will be collapsed into a single shared source of truth. Until then, the header stamp on each file is the contract.

references/scoring.md and references/examples.md are native to this repo (humanizer has no score and no diagnostic examples) and are not synced.

Supported tools

Tool	File it reads	Install
Claude Code	SKILL.md	cp -r this repo to ~/.claude/skills/authenticity-check/, or use the repo in-project
Cursor	.cursor/rules/authenticity-check.mdc	Open this repo in Cursor, or copy .cursor/rules/authenticity-check.mdc + SKILL.md + references/ into your project
Codex	AGENTS.md	Clone this repo into (or beside) your project; Codex reads AGENTS.md
Antigravity	AGENTS.md	Same as Codex: keep AGENTS.md + SKILL.md + references/ in the workspace
Gemini CLI	GEMINI.md	Keep GEMINI.md + SKILL.md + references/ in the project Gemini runs in
Pi Coder	AGENTS.md	Point Pi Coder at this repo / its AGENTS.md
OpenCode	AGENTS.md or SKILL.md	Copy the skill into OpenCode's skills directory, or keep AGENTS.md in the project
GitHub Copilot	.github/copilot-instructions.md	Copy .github/copilot-instructions.md + SKILL.md + references/ into the target repository
Windsurf	.windsurfrules	Keep .windsurfrules + SKILL.md + references/ in the project Windsurf opens
Cline	.clinerules	Keep .clinerules + SKILL.md + references/ in the workspace Cline runs in
Continue	.continue/rules/authenticity-check.md	Keep .continue/rules/authenticity-check.md + SKILL.md + references/ in the project
Zed	AGENTS.md or .continue/rules/authenticity-check.md	Zed reads AGENTS.md (already present); the Continue rule also applies if used
Aider	CONVENTIONS.md	aider --read CONVENTIONS.md, or set read: CONVENTIONS.md in .aider.conf.yml; keep SKILL.md + references/ in the repo

The Cursor rule and the frontmatter name are authenticity-check, distinct from humanizer's humanizer, so both skills can be installed side by side without collision. Every adapter points the agent at the same SKILL.md and references/, so the workflow is identical across tools. The 13-tool count includes Zed as a distinct host even though Zed reads adapter files already present (AGENTS.md or the Continue rule); no Zed-specific adapter is needed.

Usage

Ask, in plain language, whether a text is authentic, whether it reads like AI or like a person, how human a passage sounds, which parts sound machine-written, or whether your draft still sounds like you. You do not need to say "authenticity check." Oblique cues ("does this sound like a bot," "something about this feels generated") trigger it too.

For a "does this still sound like me" check, do one of:

• paste a sample of the target writing,
• name a well-known author, or
• keep a VOICE.md (schema in references/voice-matching.md) or a STYLE-GUIDE.md in the project; it is discovered automatically.

Every run returns the report: band and score, flagged spans, what was deliberately not flagged, the score basis, a caveat, and a next step. It never returns rewritten prose.

Verification

evals/evals.json holds the verification cases (an AI-heavy text, a voice-deviation check, a restraint case, a detector-evasion refusal, an oblique trigger, and a diagnose-then-"just fix it" boundary). evals/RESULTS.md records a blind verification battery: every case run as an isolated diagnosis with no access to the expected answer, plus a known-vs-non-known battery and a relocated-signature regression set. The regression set exists because the hardest case is the one that matters most: AI prose with the slop vocabulary removed but the uniform rhythm kept must still read low, while genuine careful human prose, even when formal and clean, must not be over-flagged. These files are documentation and verification only; they are not part of the runtime skill.

On short inputs (a single paragraph or two), the internal-consistency pass (Pass 3) needs at least three comparable chunks and is skipped; Step 0b plus Passes 1-2 carry the read, with the relocated-signature override holding short marker-free uniform inputs in the low band.

Scope

This skill gives an honest read of how authentically text reads as a person's work. It is not designed or tuned to defeat plagiarism checkers or AI-detection systems, and it names no detector. Requests framed as getting AI work past a graded or contractual assessment are reframed toward the honest diagnostic the skill actually serves. The diagnostic-only boundary is part of that guarantee: because the skill never rewrites and never carries a score into a transformation, it cannot be turned into a score-then-rewrite gaming loop.

Layout

SKILL.md                          orchestrator: workflow, guardrails, output contract
AGENTS.md                         cross-tool entry point (Codex, Antigravity, OpenCode, Pi Coder)
GEMINI.md                         Gemini CLI context
.cursor/rules/authenticity-check.mdc  Cursor project rule (named to not collide with humanizer)
.github/copilot-instructions.md   GitHub Copilot instructions
.windsurfrules                    Windsurf rules
.clinerules                       Cline rules
.continue/rules/authenticity-check.md  Continue / Zed rule
CONVENTIONS.md                    Aider conventions
references/tell-patterns.md       vendored, synced from humanizer: the 32-pattern catalog (Pass 1)
references/do-not-flag.md         vendored, synced from humanizer: false positives, human markers (Pass 2)
references/voice-matching.md      vendored, synced from humanizer: voice reading (Pass 4)
references/scoring.md             native: band + 0-100 rubric, internal-consistency heuristics
references/examples.md            native: four worked diagnostic runs
evals/evals.json                  verification cases (not part of the runtime skill)
evals/files/VOICE.md              voice baseline used by the voice-deviation eval

License

MIT. See LICENSE.