The NEO Fork

NEO is Hugh Howey’s word processor for authors: free, MIT licensed, and built on the idea that nothing should get between you and the next sentence. I want to fork it for Grunt Force and add a few distractions, the kind that push the next sentence along instead of pulling me away from it. This page is the plan. Nothing here is built yet.

Why NEO

It’s small enough to change. NEO is plain JavaScript in Electron with no build step: one file for the window and the disk, one for the whole interface, and a bridge between them. Books are plain files, chapters as HTML and metadata as JSON, so anything I add can read the manuscript without a database in the way. The repo even ships notes for coding agents, which makes it friendly to work on with Claude.

The rule I’m bending, carefully

Hugh’s first ground rule is that nothing interrupts a writer mid-sentence: no popups, no squiggles, no notifications while typing. I want to keep it. Every distraction in the fork waits for a pause, a full stop and a beat of silence, and shows up as a gray ghost line under the cursor, styled like the ghost lines NEO already uses for outline notes but gone the moment I start typing. They stay off until switched on, behind NEO’s own “Show More Features” switch, and they’re drawn with the CSS Highlight API the way NEO draws search results, so none of it is ever saved into a chapter file.

The sentence shape suggester

You finish a sentence and stop. A faint line appears: Try a short one. Your last five sentences all ran 18 to 26 words. That’s the whole feature. It names one of the nine shapes from my Writing Techniques notes and says why in one line. It never writes the sentence for me.

It decides in two layers.

The rhythm layer

Plain rules, instant and offline. It measures the current scene’s sentences and punctuation, using the rules from my Writing Techniques notes:

The story layer

Optional, and it reads meaning. It sends the scene so far, the chapter’s outline note and the narrator’s name to a local model on my own machine, through Ollama, the same setup that runs this site’s chat assistant, so the manuscript never leaves the computer. It picks the shape that fits what the scene is doing: a delay before a reveal, a reversal when a promise sours, a list when a character sizes up a room.

Narrator leans

Each narrator gets a lean, from the voices in my style guide: Owen long and anxious, so pile-ons and long sentences; Kevin dry and clinical, so short sentences and contrasts; Andre quick and guarded, so short ones and lists; Mateo warm, so building sentences and echoes. Those leans are my guesses. Once each narrator has a voice sheet (the eight parts in Writing Techniques), the leans come from the sheet, and the fork can measure whether a chapter matches its narrator: sentence length, passive voice, hedges, the words they repeat.

Revision passes

The suggester is the only thing that talks while I draft. Everything else is a pass I ask for, the way NEO already makes spellcheck something you invoke. Each pass marks the text with the CSS Highlight API and a margin list, and none of them rewrites a word.

What each pass flags

• Words to cut: the table from Writing Techniques, counted per thousand words, with dialogue left out of the hedge count.
• Point of view: since every chapter has one narrator, flag the narrator describing their own face, thought tags like “she thought”, and filter verbs after the narrator’s name.
• Clutter: three prepositional phrases in a row, body language that repeats itself (“nodded her head”), and sentences that only move someone between rooms.
• Description: a long block with no dialogue at a chapter’s start, paragraphs that only use sight, and a first description of someone that leads with eye and hair colour.
• Dialogue: an action beat on every line, or long runs with none, and a voice check that compares each speaker’s words, sentence length and pet phrases.
• Time and cast: a chapter whose first two sentences don’t say when we are, and paragraphs that introduce more than two new names.

Outline cards that ask questions

NEO’s outline already lays every chapter and section out as an index card. The fork would add a few optional fields to each card, from the scene methods in my notes: what the scene’s question is, its forwarding device (new information, a decision not acted on yet, or a complication), whether it ends on “yes, but” or “no, and”, its one tentpole moment and its one revelation. An empty field is a quiet gray prompt, never a warning.

A book view would chart what a card can’t see: chapter lengths in a row, dialogue share per chapter (two or three heavy ones in a row drag), runs of fast chapters with no breather, and which narrator has gone longest without a chapter.

How it would fit in

• The rhythm rules and the passes live in new files of pure functions, so they’re tested with Node’s own test runner the way NEO tests its spellcheck and dashes.
• The suggester hooks into the typing section of NEO’s interface file and borrows the outline notes’ ghost styling for its line.
• The model call goes through NEO’s bridge to the main process, since NEO’s window never touches the disk or the network itself.
• New code stays in its own files where possible, so pulling Hugh’s updates into the fork stays easy.

Open questions

What counts as a pause: a full stop, an Enter, or two seconds of nothing? Whether the suggestion should stay until dismissed or fade on the next keystroke. Which local model is small enough to answer before I’ve started the next sentence anyway. And whether the fork needs a name of its own.

Sources

NEO’s features come from Hugh’s video and the repo’s README; its rules, structure and More Features switch from the repo’s contributing guide and agent notes, read at version 1.6.1 in October 2026. The rules the suggester and the passes would check come from the Bookfox videos digested in my Writing Techniques notes, where each one has its source card. The plan is mine.