Voiceprint

One entry point for prose authenticity. The user gives it text and gets back a result: the original, a humanized rewrite, an honest read of what still looks machine-touched, and a plain statement that the rest is a human's call. Voiceprint composes two existing prose skills in a fixed order, exactly once. It also applies one narrow, deterministic Unicode hygiene operation at the start of the existing transformation stage. That local operation cleans the working copy without adding another prose rewrite.

The user never invokes humanizer or authenticity-check directly through this skill. Voiceprint orchestrates both internally from vendored copies. Those two skills remain canonical and independently usable in their own repos; this one keeps their prose methods intact.

When to use this

Use it when someone wants the whole loop closed in one step: fix the draft and tell them how it reads now. Triggers include "clean this up and verify it," "make this sound human and tell me how it scored," "de-slop this then check it," "voiceprint this," or any phrasing that asks for the rewrite and the read-back together. Apply it even when they do not say "voiceprint."

Do not use it for a one-sided request. A pure "rewrite this so it sounds human," with no interest in a score, is the standalone humanizer skill. A pure "is this AI / how human does this read," with no rewrite wanted, is the standalone authenticity-check skill. Voiceprint is specifically the union: diagnose, fix once, re-read, report. If the user only wants one half, serve just that half and stop: prefer the standalone skill when it is installed, and otherwise follow the matching vendored copy directly (vendor/humanizer/ for a rewrite-only request, vendor/authenticity-check/ for a score-only request). Do not run a half-empty voiceprint pass: a rewrite-only request gets no authenticity read, and a score-only request gets no rewrite.

Core principle: one principled pass, never a loop

This is the rule the entire design exists to enforce, and it does not relax.

Voiceprint runs the diagnostic, applies the rewrite once, and runs the diagnostic again only to report the residual. The second read informs the user. It never feeds back into another rewrite. There is no target score, no threshold to chase, no "the residual is still high so let me try again." A tool that scores text and then rewrites it to raise its own score is a detector-gaming loop that optimizes prose against a metric instead of improving it for a reader. Both upstream skills refuse that loop on purpose; keeping diagnosis and transformation as a single human-judged pass, not an automated cycle, is the whole reason this product is shaped the way it is.

What remains after the one pass is reported, not resolved. Resolving it is a human's judgment, and the output says so explicitly.

The pass

Run these three steps in order, once each. Do not improvise a shortcut, and do not repeat a step.

Step 1: Diagnose (once)

Capture the submitted text as an immutable original. Read vendor/authenticity-check/SKILL.md and follow it exactly on that original text, including the reference files it points to under vendor/authenticity-check/references/. Produce its full authenticity report (band, score, provenance signals, flagged spans, what reads as human, score basis, caveats). This is the before read. Its provenance preflight only inspects and reports the original's Unicode carriers; nothing is cleaned in this step. The Before section must retain the submitted text byte-for-byte, and Step 1 must receive the same code-point sequence. Carry no target score out of it.

Step 2: Clean the working copy and apply humanizer's fixes (once)

At the start of this step, run the immutable original through python3 scripts/text_hygiene.py clean --stats exactly once, resolving the script relative to this SKILL.md. Python 3.11 or newer is required. For pasted text, prefer the command's standard input and pass the original bytes through the host process-input facility. Never interpolate pasted text into a shell command, including a heredoc: a line in the draft that matches the delimiter would end it and run the rest of the draft as commands. A user-provided path remains a read-only source.

If the host cannot supply standard input and a temporary file is unavoidable, use its secure-temp facility, require owner-only permissions, and guarantee cleanup when the command succeeds or fails. Never rewrite the source file in place. Standard output is the cleaned working copy; standard error is the deterministic JSON manifest. Input is capped at 4 MiB, which comfortably exceeds the 100,000-code-point product requirement.

If the helper exits with any status other than 0, stop before humanizer runs. Exit 2 means the input is too large, is not valid UTF-8, cannot be read, or could not be processed or written; any other status, such as a process the host stopped, is a failure too. State that the original was not changed and do not claim cleanup succeeded.

Pass only the cleaned working copy to one invocation of vendor/humanizer/SKILL.md and follow it exactly, including the reference files under vendor/humanizer/references/. Run its method as written (voice discovery, density pre-check, text-hygiene preflight, the multi-pass workflow, the meaning check). Apply it a single time. This produces the after text. Humanizer's own text-hygiene preflight and final recheck run inside this one invocation, on the already cleaned working copy. They are part of humanizer's method, not a second cleanup stage, and the helper's manifest remains the authoritative hygiene record. Do not loop the rewrite, and do not let the Step 1 report set a goal for it; the rewrite's only job is to apply humanizer's known fixes once, faithfully, with humanizer's own restraint and anti-fabrication guards intact. Do not add candidate generation, statistical watermark rewriting, or another cleanup stage.

Step 3: Re-diagnose for residual only (once)

Read vendor/authenticity-check/SKILL.md again and run it once on the after text, as a fresh, cold diagnosis with no carry-over from Step 1 and no score target. Its output here is the residual: what still reads as machine-touched after one honest rewrite, plus any provenance signal still present in the after text. Report it. Do not act on it. If the residual read is poor, or a residual provenance signal remains, that is information for the user, not a trigger to clean or rewrite again. The pass is over.

Hard rule: no iteration

The re-check never drives further rewriting. Not when the residual score is low, not when spans remain, not "just one more quick fix." One diagnose, one rewrite, one re-diagnose, then stop and report. If you are tempted to run Step 2 a second time because Step 3 came back unflattering, that temptation is exactly the failure mode this skill forbids. Stop.

Output contract

Always deliver this exact structure. Never silently rewrite in place.

## Voiceprint result

### Before
[the original text, verbatim]

### Authenticity read (before)
[the Step 1 authenticity-check report: band + score, provenance signals,
flagged spans, reads-as-human, score basis, caveats]

### After
[the Step 2 humanized text, the primary artifact]

### What changed
[first report text hygiene as detected, removed, normalized, and deliberately
preserved counts, then include humanizer's own "What changed" / "Deliberately
left alone" / "Meaning check" sections, unaltered]

### Residual (after one pass)
[the Step 3 re-diagnosis: band + score, residual provenance signals, the
spans that still read as machine-touched. This is a report, not a to-do
list.]

### What remains is a human's call
One short paragraph stating plainly that voiceprint ran exactly one pass, that
the residual above was deliberately not rewritten further, and that closing
any remaining gap is human judgment, not another automated rewrite. Name the
single biggest residual factor so the human knows where to look.

The "Residual" and "What remains is a human's call" sections are not decoration. They are the forcing functions that make the no-loop rule visible to the user and impossible to quietly skip.

This contract supersedes the vendored skills' own output wrappers. Fold humanizer's Voice:, Density:, and Text hygiene: header lines into "What changed" rather than printing them separately. Do not emit either vendored skill's standalone "Next step" inside the pass: humanizer's offers to write the rewrite into a file, and authenticity-check's sends the user back to a rewrite and a fresh read, which inside the residual would invite exactly the loop this skill forbids. The "After" text above is the artifact. If the user gave a file path and wants it persisted, offer that once, after the full contract is delivered, never as a silent in-place rewrite. The pass itself never edits files. Because it reads untrusted text, this skill pre-approves only read-only tools: running the hygiene helper, and writing a temporary input file when the host has no standard-input facility, go through the host's normal permission prompt.

Inside What changed, identify hygiene findings only by their observable code point, Unicode name, action, count, and up to 10 zero-based code-point offsets. Include the helper's preservation reason for every deliberately preserved character. If the manifest's findings list is empty, state explicitly that detected, removed, normalized, and deliberately preserved counts are all zero. Report the manifest counts before humanizer's Text hygiene: line. That line describes the already cleaned working copy, so it can report nothing suspicious even when the helper removed characters. These details stay inside What changed; do not add a seventh top-level section.

Scope and intended use

Voiceprint exists to improve prose quality and to help a writer's own work sound authentically like them, then to give an honest read of what is left. It is not designed or tuned to defeat plagiarism checkers or AI-detection systems, and it deliberately names and optimizes against none. The residual read is a heuristic, not proof of authorship, and a low residual is not a guarantee a detector is fooled, because fooling a detector is explicitly not what this tool does. If a request is framed as passing AI-written work off as a person's own for a graded or contractual assessment, do not adopt that framing; offer the quality-and-voice improvement and the honest read instead, which is what voiceprint actually does well. This restates the boundary that the vendored humanizer and authenticity-check skills already hold; voiceprint inherits it without exception.

Text hygiene reports deterministic Unicode observations only. A removed control or normalized space does not reveal who created the text, whether a watermark was present, or how an external detector will classify the result. The same holds for the provenance signals in either authenticity read. Do not claim vendor provenance, watermark removal, detector-signal removal, or a passing score. The hygiene helper does not inspect file metadata, document properties, images, audio, or video, and it does not perform NFKC or confusable letter conversion.

Composition and sync obligation

humanizer and authenticity-check are canonical upstream and independently usable. Voiceprint vendors copies of both skills, plus the shared detection criteria, under vendor/. The sync is one-directional:

• Canonical upstream for detection and rewrite criteria is the humanizer repo.
• Canonical upstream for scoring logic is the authenticity-check repo.
• Voiceprint never edits vendored content. Any criteria or logic change goes upstream first, then syncs down via scripts/sync-upstream.
• Every vendored file carries a header stamping its true upstream repo, the source commit, and a "synced copy, do not edit here, edit upstream" notice.

The full sync procedure is in README.md. Editing anything under vendor/ by hand is the one move that breaks this product.

Reference files

Read these on demand, during the pass, not upfront:

• vendor/authenticity-check/SKILL.md in Step 1 and Step 3, with its references/ (scoring, provenance-signals, tell-patterns, do-not-flag, voice-matching, examples).
• vendor/humanizer/SKILL.md in Step 2, with its references/ (tell-patterns, do-not-flag, voice-matching, text-hygiene, examples).