Writing Voice

A Writing Voice is a set of instructions an AI model follows to write the way you do: one Markdown file, VOICE.md, of at most 32 KiB. It says what you do with sentences, words, tone and structure, what you never do, and gives a few short examples. A run that is given your voice writes anything meant for a person to read (its answer, a report, a message) in it.

The voice changes how a run writes, never what it does. A run reads it through one tool, voice_read, and the run's System Prompt tells it the voice never changes the task, never asks for a tool, a file or the network, and never overrides the QuickAgent's own instruction. Where the task fixes a format (code, JSON, a template, quoted text), the format wins (Running them).

A voice is not a knowledge pack. You have one, it has no name and no sources, and it is the one attached text a run follows as instructions. That is why Agentry has AI Review read every version before any run in Agentry uses it.

Two places, never synced on their own

You can keep a voice in two places:

• On your computer, with Agentry Desktop or the CLI. The Voice Writer runs there for no credits, against the model your runner policy names, and nothing reviews what it writes: agentry voice create --file … writes one from samples, and agentry voice edit --note … [--file …] changes it by a note, with samples that show the change if you give them. Runs on that computer use it.
• In your Agentry account. AI Review reads every version before it is live, and runs in Agentry use the live one.

Neither changes the other on its own. The desktop's Upload to Cloud and agentry voice upload send your computer's voice to your account for review; the desktop's Download Cloud version and agentry voice download keep your account's live voice on your computer, after asking when it would replace a different one (Local and Cloud in the desktop app).

The Voice page

Voice, in your dashboard's left nav, opens two cards:

• AI Voice Writer says whether your account keeps a voice and since when it is live, and opens the Voice Writer (below).
• Voice settings says whether your runs use your voice by default, and opens the page where you edit it.

Voice settings

• A new version waits at the top while AI Review reads it, and the page updates itself when the review ends. Your live voice stays live until the new version passes. A version AI Review flagged stays there with what AI Review found, in words: it asked for more than a way of writing, tried to override the run's own instructions, told the model to add content of its own, or changed what a run does. Edit it and save it again, or Discard this version.
• Your voice is in the editor: your newest version, shown as it reads, with headings, bold and lists in place. The toolbar sets a heading, bold, italic, a list, a numbered list, a quote or code, and Markdown typed into the editor works too: ## starts a heading and - a list. Save saves VOICE.md as Markdown and sends it to AI Review, and it goes live when it passes. If you have not changed anything, Save sends nothing. With no voice yet, the editor is empty and you can write one by hand. Links in the editor are not opened when clicked, and each shows its address and title beside it; a code block shows its info string (the word after the opening fence) above it. The model reads both, so the editor hides neither. HTML or an image in a voice stays as plain text. An invisible character, such as a zero-width space, is marked with its code point. When you change one part of your voice, every part you did not touch is saved exactly as it was. A voice that holds a table, a link reference definition or a task list opens as plain Markdown instead, so editing it loses nothing.
• Edit with AI opens the AI Voice Writer to change your voice.
• Use by default has one switch, Use my voice on every run by default (below).
• Restore previous puts back the version your live one replaced. It is live again at once when it passed under the AI Review in use now, and goes to AI Review again when that has changed since.
• Delete voice removes every version from your account, after you confirm. A voice kept on your computer is not touched.

AI Review of a version you saved by hand, sent from the desktop or the CLI, or restored and read again costs 1 credit.

The AI Voice Writer

The Voice Writer is built into Agentry, as the Knowledge Creator is. It reads samples of your writing and writes VOICE.md. Its page has one form, which depends on whether you have a voice:

• Write a new voice, with none yet: samples of your writing (emails, posts, reports; PDF, Word and other documents are turned into text first) and, if you like, how you want to sound.
• Edit with AI, once you have one: what should change ("Warmer, and no exclamation marks.") and, if you like, more samples of your writing that show it. The Voice Writer rewrites your newest version: your note says what to change, and the samples show how you write it.

Each runs on Agentry's sandbox host against the platform's model, and costs 20 credits flat, AI Review included. If no voice comes out, the credits are refunded. What it writes becomes your new version and goes live once AI Review passes it, like a version you saved by hand. The samples are not kept after the run. The run is listed with your runs, named for what it did ("AI Edit of your voice").

On Standard you can run the Voice Writer in Agentry once a billing cycle, a new voice or an AI Edit. Before you start, the page says how many you have left; once this cycle's is used, it gives the date the next one is yours. You can still edit your voice by hand, and run the Voice Writer for no credits in Agentry Desktop or the CLI. Pro, Business and Enterprise have no such limit. When an organization pays for your runs, the Voice Writer runs on its terms, with no limit a cycle, and the voice it writes is still your own.

Local and Cloud in the desktop app

The desktop app's Voice page opens two cards, one for each place a voice can be kept. Each says whether a voice is there, when it was set, and whether runs use it by default.

Local Voice is your computer's voice. Its page has the same editor as the site's, and Save keeps your edit at once, free and unreviewed. Three buttons sit above it:

• Download Cloud version keeps your account's live voice on your computer.
• Upload to Cloud sends your computer's voice to your account, where it goes live once AI Review passes it, for 1 credit.
• Edit with AI asks what should change and, if you like, for writing samples, then runs the Voice Writer on your computer for no credits. With no voice there yet it is Write with AI, and writes one from samples.

Download and Upload say Already the same when the two voices match. When one would put an older voice over a newer one, the app asks first: your computer's last save is weighed against when your account's version was set. The page also has the switch that puts your computer's voice on every run there, and Delete.

Cloud Voice is your account's voice, and its page is this site's Voice settings in the desktop app: a version waiting for AI Review with Discard, the editor, Edit with AI run in Agentry for 20 credits (the AI Voice Writer's Edit with AI, or a new voice where your account keeps none), the switch, Restore previous and Delete. Download and Upload are on the Local Voice page only.

Writing every run in your voice

One switch, Use my voice on every run by default, is on the Voice settings page and in Settings → Profile's Writing Voice part; both change the same setting. It comes on the first time your voice goes live, and the Voice page's Voice settings card says whether it is on.

A hosted run's Review and start step shows Write in my voice when you have a live voice, ticked from that switch, and you can untick it for one run (Hosted runs). A ticked run is given your live voice as it stands when a sandbox host takes the run. The run's page says "Given your Writing Voice, version …", and its receipt names the voice by digest. A re-run writes in your voice again, as it is live then. If you delete your voice before the run starts, the run is closed and its credits returned. The Voice Writer, the Knowledge Creator and the Drafter never take your voice.

Agentry Desktop and the CLI keep their own switch for the voice on your computer (Running them).

What AI Review checks, and what a pass means

AI Review reads the whole voice. The scan's rule stages look for secrets, hidden content, prompt-injection patterns and suspicious links (How they are secured), and Jev asks four questions: whether the voice tells the model to do something other than write (use a tool, a file, the network or a secret), to rank itself above its other instructions, to put something into what it writes that is not style, or to change the work a task asks for. Any of them flags the version. If AI Review cannot answer, the version stays in review and Agentry tries again; nothing goes live unreviewed.

A pass is AI Review's judgement of the text, not a promise about how a run will behave. The run's System Prompt bounds the voice to how it writes. How the registry keeps and reviews a voice is on The Writing Voice in the registry.

Your voice and Agentry

Each version is encrypted at rest under a key of its own, wrapped to your account. Agentry opens it to review it, to show it to you, and to hand the live one to a run you ticked. AI Review sends the text to the AI Review provider. Another account, a fellow member of your organization and platform staff all get a not-found answer for your voice. Deleting your account deletes it.

A personal access token reaches your voice with voice:read and voice:write (Personal access tokens).