Gavin JuddIndex

Work

Voicework

I built Voicework in October 2026 as a local writing application. It drafts text for a particular audience, purpose and channel, then lets me edit the result and asks which of my corrections it should remember. The design question is narrow: how can it remember a correction that helps in one situation without quietly making it a rule for everything I write?

Saving an edit is its own step

An edit might fix a fact, adjust the tone or make one sentence work in that particular document. Those are different reasons to change a draft, but saving every edit as a permanent description of someone’s voice would lose those distinctions.

That’s why remembering a correction is a separate action. Before saving anything, the application shows the user the original wording, the revision and a plain-language takeaway. They can keep the lesson tied to a specific audience, intent, relationship or stakes, or deliberately apply it across a whole channel. Changes to facts or meaning stay with the draft; they don’t become style rules.

The rest of the data follows the same principle: authored samples, remembered corrections, direct language lessons, background information and evaluation votes are stored as different kinds of evidence. A correction to AI-generated text can still guide later drafts while remaining correction feedback, separate from authored samples.

Undo

Every remembered correction keeps its provenance, along with an undo action that stays available. Undoing it disables the lesson that came from that correction while leaving unrelated lessons alone.

If a lesson has been edited since it was saved, the application sends the user to review the current lesson instead of letting undo overwrite the newer version. The original correction stays in the local history, and undo doesn’t claim to erase earlier exports or writing that’s already been generated.

Checking one correction at a time

There are two ways to compare drafts. One compares a calibrated draft with a baseline, while the other asks a smaller question: does this particular correction help?

For the smaller check, the brief, current examples and other preferences stay the same across both drafts; the selected correction is the only input that changes. Both the generation order and the display labels are randomized. The user gives separate ratings for voice, meaning and expected editing effort, and that vote never trains the profile or switches a rule on or off. The saved record includes the model, the controls and fingerprints of the inputs, but it describes only that comparison under those conditions. I don’t treat it as an accuracy score.

Whole documents

For longer pieces, Voicework freezes the selected profile before rewriting the document in bounded sections. The result includes editable text, separate content and voice checks, structure checks, and hashes of both the source and the draft. Editing the output marks those checks as stale, so rechecking is a separate step. If a review is incomplete, the application reports it that way rather than calling it clean.

Another tool can use the same workflow through a JSON command-line harness: it sends a document and context and gets back the draft, its checks and the context that was selected.

How it’s built

Voicework is served from the user’s own machine and uses Python, JavaScript and SQLite, with no third-party packages. It personalizes drafts by pulling context-matched examples and saved feedback into the prompt, without a fine-tuned model. Generation can run through Codex or, since October 2026, through Claude. For Claude, it can use either a Claude Code login or the Anthropic API. The selected provider receives the chosen examples, lessons and brief, so storing the data locally doesn’t mean generation happens offline.

AI coding agents did some of the implementation and review, including document generation and its checks. Voicework runs for one person on one computer, so there’s no hosted version.

The source code is available under the MIT License.