copydesk

Copydesk

Copydesk turns findings into publishable writing in your own voice. You feed it your raw material (an investigation's findings, research notes, a transcript, an outline) and it drafts in your own voice, then runs the draft past two reviewers before you ever see it: one that hunts AI tells and voice drift, one that checks whether the writing actually lands. The output comes back clean, with a short advisory table you can accept or reject row by row.

It learns the voice from samples of your own writing, and it keeps refining from your edits. Every time you fix a draft by hand, Copydesk can study what you changed and sharpen its rules so the next draft needs fewer fixes. Your writing, your registers, and everything the learning loop accumulates stay on your machine, inside your own Claude Code session.

How it works

Your raw material → register select (pick the voice that fits) → draft in your voice → a dual review gate runs prose-review and craft-review in parallel → hard fails are fixed silently and everything else comes back as an advisory table → your final draft. When you edit it by hand, the learning loop studies the change and sharpens the rules for next time.

What you can do with it

• Write up findings in your own voice. Hand Copydesk an investigation's findings and ask it to write the piece. It activates automatically whenever you ask Claude to write text for an audience (a blog post, an article, advocacy copy, an email, a newsletter), drafts in the register that fits, and runs the review gate before you see anything.
• Keep separate voices for separate work. Configure one register for advocacy copy and another for personal essays. Copydesk detects which fits the context and loads that voice profile as the primary instruction. If it's ambiguous, it asks which register to use.
• Catch AI tells before they reach the page. The gate fixes em dashes, the "this isn't X, this is Y" construction, and the ChatGPT-isms silently, and surfaces everything else (voice drift, structural monotony, dead transitions) as advisory rows you decide on.
• Sharpen the voice from your edits. Edit a draft by hand, then run /copydesk:learn to study what you changed. It diffs the pipeline output against your hand edits and proposes bounded improvements to your register, the skill rules, or the review agents.
• Build the voice once from your own samples. Run /copydesk:init: it sets up your data directory and walks you through a two-pass extraction over 10-20 samples of your writing, producing a register that captures your vocabulary, sentence structure, rhetorical moves, and voice qualities. Run it again any time to add another register.

Why it's useful

Most AI writing reads like AI writing. It puffs up significance, leans on the same dozen transition words, and forces every idea into a group of three. A reader can feel it even when they can't name it. Copydesk's job is to make the writing sound like a specific person wrote it, because a specific person did the thinking, and then to defend that voice with a review pass that knows exactly what machine writing looks like.

The two reviewers pull in deliberately different directions. One is tuned for precision: it fires rarely and is meant to be right, catching banned phrases and voice drift. The other casts wide for craft opportunities you'd otherwise miss (a pattern you described but never named, an ending that summarizes instead of landing). Most of what it surfaces gets rejected, and that's the design working. You stay in control of every judgment call. Hard fails get fixed for you; everything else is a suggestion you can wave off.

The learning loop compounds. The initial extraction gets you a usable voice. Your edits make it precise. Over a handful of pieces, the drafts tend to need fewer corrections, because the system has watched what you actually do and adjusted its rules to match, with a held-out gate making sure each change is a real improvement and not just noise.

Quick start

Install it from the Fieldwork marketplace:

/plugin marketplace add TimSimpsonJr/fieldwork-plugins
/plugin install copydesk@fieldwork

Build your voice once. Run /copydesk:init (about 30 minutes, using Claude Sonnet). It creates your data directory at ~/.claude/data/copydesk/ and walks you through gathering writing samples and two extraction passes, leaving you with a register Copydesk draws on automatically.