Inspiration
I edit book manuscripts, and every project ends up the same way: a folder holding a dozen versions of the same book, with no clear record of which one is current. Names and terms drift between drafts, citation formats change halfway through, and the style sheet has to be rebuilt by hand on every pass. Authors are also protective of unpublished work, so sending a full manuscript to a cloud AI tool is something many of them refuse to do. A NAS that can read and reason over files locally is a natural home for this job.
What it does
Manuscript Desk is an AI NAS Skill for authors and editors. The user points it at a manuscript folder, and it finds every draft, groups them into one book, identifies the latest version, and produces three working documents: a draft index, a style sheet covering names, terms, spelling and citation format, and a consistency report that flags problems such as a name spelled two ways, a skipped chapter number, or a passage duplicated across chapters. It reports and never rewrites, it treats manuscripts as read-only, and it asks for confirmation before saving anything.
How we built it
The Skill is a Markdown package built to the UGOS Skill standard: a SKILL.md that defines a seven-step workflow, plus reference files for the style sheet format, the consistency checks and disclosures. Each step maps to one NAS capability: Files to find and tag drafts, Documents to read, extract and reconcile versions, and Knowledge to save the results. I used Claude through Claude Code to help draft and review the Markdown, while the editorial rules come from my own experience of editing books.
Challenges we ran into
The hardest part was designing for a long manuscript within the limits of a local model, which led to processing chapter by chapter and capping quoted evidence at fifteen words per finding. The other challenge was restraint: keeping the Skill to reporting only, so that an editor can trust it will never alter an author's words.
Accomplishments that we're proud of
Every step in the workflow maps to a single NAS capability, every write waits for the user's confirmation, and every path where nothing is found stops cleanly instead of guessing. The consistency checks are also the ones a working editor actually runs, ranked by how much damage each kind of error does in a published book.
What we learned
Writing a Skill is closer to writing a procedure for a careful colleague than writing software. Each step needs a clear capability behind it, a clear stopping point when nothing is found, and a clear moment where the user is asked before anything is written.
What's next for Manuscript Desk
The package has not yet been run on a UGREEN AI NAS. The next step is to test it on real hardware with three cases: a folder of twelve drafts, a draft seeded with known inconsistencies, and a folder that mixes two different books. After that, I would like to add support for comparing an editor's marked-up copy against the author's revision.
Built With
- agent-skills
- markdown
- ugos
- ugreen-ai-nas
