The PracticeThe Practice The PracticeOperating infrastructure

The Marketplace·The Voice Sweep

The Voice Sweep

Catches the machine tells in your writing and returns it to your own voice.

What it is

An editorial repair for drafts written with AI help. It finds the patterns that make prose read as generated, repairs them, keeps every fact, and shows you every change.

Rung

Instrument, one working tool for one job.

Price

$212

Each product is sold as the version delivered. We do not promise updates, fixes or support after delivery. When we release a new major version of a product, half of what you paid for the version you hold counts toward the new one, with no time limit.

Request early access

Or get the Letters, twice a month by email.

Format

16 files:

  • the skill and its instructions;
  • a guardrails file;
  • five reference files, a 30-signal checklist among them;
  • a metrics script;
  • a three-file worked demo;
  • a PDF guide;
  • a README and a disclaimer.
Who it fits

Anyone who drafts with Claude or ChatGPT and sends the result under their own name. Solo practitioners and small teams, in any field.

Version

1.0.1, August 29, 2026. Writing signals last checked August 24, 2026.

Requirements

Claude, full strength: Any plan, Free included. It installs as a Claude Project. The metrics script runs where code execution is on; where it is off, Claude estimates the numbers and says it is estimating. Claude Free holds up to 5 Projects. Claude Pro is $20 a month, or $17 a month billed annually. (claude.com/pricing, as of October 29, 2026.) ChatGPT, compressed: A ChatGPT Project holding four files: three reference files and your voice profile. It keeps the three tiers and the report. It loses the metrics script. Four files fit within the Free plan's limit of 5 files per Project (OpenAI help center, as of October 29, 2026). Install: you install it yourself, in your own account.

Built and tested on Claude and ChatGPT. Everything is plain files you keep, so you can use it in other assistants. We have not tested those.

What's included

  1. The skill, ready to upload to Claude.
  2. The instructions: calibrate to the writer, take in the draft, run three tiers of checks, repair, then report.
  3. A guardrails file the sweep reads first on every run, setting the boundaries it keeps.
  4. A 30-signal checklist in three tiers, each signal weighted by the strength of its evidence.
  5. A calibration guide that builds your voice profile from your own writing, so your habits are protected rather than flagged.
  6. A prescription file of drafting habits, for writing fresh to the same standard.
  7. Worked examples of repairs.
  8. Install guides for Claude and ChatGPT, each stating what that platform gives up.
  9. A metrics script for sentence rhythm and paragraph variance.
  10. A worked demo: a sample draft, the swept draft and the full report.
  11. A PDF guide to the whole pack, and a disclaimer.

Every run returns the repaired draft and a report:

  • metrics before and after;
  • every repair, quoted with its tier;
  • the lines it kept on purpose, and why;
  • questions only you can answer;
  • two or three habits that would have prevented the draft's biggest flags.

Sample output

Sample: An excerpt from the demo report bundled with the product, run on a sample market-update email.

Metrics (before / after): Burstiness 0.60 to 0.51 with the fragment drumbeat removed and real long-short variation added; the before number was inflated by four staccato fragments, which is rhythm noise, not rhythm. Paragraph variance 0.50 to 0.73 (paragraphs now carry different weights). Words 164 to 208, and every added word is a fact you supplied. Tier 2 hits 8 to 0. Glue phrases 2 to 0. Antithesis kickers 2 to 0 stacked (one survives inside a sentence: "a working market, not a frozen one").

Repairs by tier (Tier 1, architecture; excerpt):

  • Fungible closer: "unlock your real estate goals and elevate your journey" could end any agent's email in America. Replaced with an offer only you can make: numbers for their street, not the metro average.
  • Evidence mode: unnamed authority ("Experts say") cut entirely; every claim now has your source or your incident behind it.
  • Ranked judgment added: the email now tells rate-watchers something they may not want to hear (the dip they are waiting for is the same one everyone is waiting for), which is what makes it read as a person with a view.

Protected lines: Your subject-line habit of leading with specifics; the direct reply-or-call close; first person throughout. No calibration profile exists yet for you, so Tier 3 caps ran at defaults.

Prescriptions (excerpt): Advance something in every paragraph. Paragraphs two and three both said the market is changing without a fact, name, number, or consequence; one of your three numbers in each would have stopped the circling before it started.

What it won't do

  • It makes no claim about AI detectors, and it does not test drafts against them.
  • It never supplies a fact, statistic, source or story you did not give it. A missing number comes back marked EVIDENCE NEEDED.
  • It does not edit quotations, citations, contract or disclosure language, or legal disclaimers. It flags them and leaves the words alone.
  • On listing copy it reads voice only. It is not a fair-housing read. The Fair Housing Sweep is the product that flags fair-housing language for you and your broker to review.
  • It does not write new content under the name of a sweep. New writing is a separate, stated task.

Skip it if you want a tool that writes for you from nothing, or a score that says a draft will get past a detector.

Trade-offs

The Guardrails PackThe Voice SweepThe Written Practice
PriceFree$212$451 ($530 bought separately)
What it doesSets your assistant's standing rules for truth, balance and roleRepairs a draft and shows every changeThe Voice Sweep, the Clarity Audit and The Calibration Field Guide together
What you give upAny repair of a finished draftA check that the words say what you meanNothing in this product; $239 more
What you addLine-by-line repair with a reportThe Clarity Audit's meaning check, and a guide to setting your assistant's judgment

Every rung, side by side: What it costs.

Common mistakes

  • Sweeping before you build a voice profile. With no profile, nothing is protected, and the sweep can sand off your voice along with the residue. Build the profile once.
  • Protecting a signature habit without an example. A protected device needs a quoted line from your own writing from before you used AI.
  • Keeping a run that supplied a statistic. That breaks the guardrails. Discard the run and point the sweep back to its guardrails file.
  • Letting long ChatGPT sessions drift. When the report loses its structure, paste the structure back in. On ChatGPT that is a normal working habit.

Questions

Will a swept draft get past an AI detector?

The Voice Sweep makes no claim about detectors, and it will tell you so if you ask. It aims for writing that reads as yours because it carries what generation cannot supply: your incidents, your sources and your judgment.

Will it change my facts?

It is built not to. Facts, numbers, names, dates, quotations and citations are meant to stay as written, and where a repair would touch one, the sweep flags it instead of editing it. Every change is quoted in the report. Read the report before you send.

Which platforms does it run on?

Claude, at full strength, and ChatGPT, in a compressed edition without the metrics script. Those are the two platforms it is tested on, and each install guide says what that platform gives up.

Do I need a paid AI plan?

Not on Claude: it runs as a Project on any Claude plan, Free included. On ChatGPT it needs a Project that holds four files, within the Free plan's limit of 5 files per Project.

Where does my text go?

It stays in your own Claude or ChatGPT account. The Practice hosts nothing and receives nothing; your AI provider's own terms govern what it does with your text.

Can my team share one purchase?

No. The license is personal to you, for your own work, including work for an organization that employs you. A team that wants it for several people should ask us about that.

What is the refund policy?

A full refund within 14 days of purchase, no questions asked, once per product. Write to jake@runthepractice.com. A refund ends your license to that product.

Will it keep up as AI writing changes?

The checklist is dated on purpose: its signals were last checked August 24, 2026. The product is sold as the version delivered: Claude and ChatGPT change without notice, and no updates or fixes are promised. Half of what you paid counts toward a new major version, with no time limit, and a revision never ends a license you already hold.

Fine print

Prices are in U.S. dollars. Sold by Stoa LLC. Payments are processed by Stripe. Sales tax is calculated at checkout where it applies. Available to customers in the United States only. A purchase is a license for your own professional use, not ownership, and what you write with it is yours. One-time purchase. No subscription. These are tools, not legal advice, and buying them makes you no one's client.